The Second International Conference on Digital ...

Conference Title

The Second International Conference on Digital Information Processing, Data Mining, and Wireless Communications (DIPDMWC2015)

Conference Dates

December 16-18, 2015 Conference Venue

Islamic Azad University, UAE Branch, Dubai, United Arab Emirates ISBN

978-1-941968-26-0 ©2015 SDIWC

Published by

The Society of Digital Information and Wireless Communications (SDIWC) Wilmington, New Castle, DE 19801, USA www.sdiwc.net

Table of Contents

Increasing the Target Prediction Accuracy of MicroRNA Based on Combination of Prediction Algorithms ………………………………………………………………………………………………………………… 1 Clinical Prediction of Teeth Periapical Lesion based on Machine Learning Techniques ………………. 9 Extraction of Causalities and Rules Involved in Wear of Machinery from Lubricating Oil Analysis Data ………………………………………………………………………………………………………………………… 16 ‘Fuzzy’ vs ‘Non-Fuzzy’ Classification in Big Data …………………………………………………………………………. 23 On Definition of Automatic Text Summarization ……………………………………………………………………….. 33 Predictive Analytics of Student Graduation Using Logistic Regression and Decision Tree Algorithm …………………………………………………………………………………………………………………………… 41 Sentiment Analysis of Social Media for Evaluating Universities …………………………………………….……. 49 Adopting Games Development and Visual Curriculum Design (VCD) Framework for Connected eLearning ……………………………………………………………………………………………………………. 63 Brain - Computer Interface for Communication via Record Electrophysiological Signals ………….… 69 Graph-type Classification Based on Artificial Neural Networks and Wavelet Coefficients …………… 77 Development of an Intuitive Data Transfer Application for 3D Video Communication System in Synchronized AR Space ………………………………………………………………………………………………. 86 A Question Answering System being able to Infer based on Definition and Acquire of Knowledge in English Text of the Written Language ……………………………………………………………..…… 93 Abstraction Method of Words Based on Relationships between Subjects, Objects and Predicates …………………………………………………………………………………………………………………………… 100 Identification and Evaluation of Keyphrases: Fuzzy Set based Scoring and Turing Test Inspired Evaluation ………………………………………………………………………………………………………………………………….. 107 Effectiveness of Scientific Methods in Detecting Phishing Websites in Jordan ………………………..…. 118

Efficient Adaptive Tree-based Protocol for Wireless Sensor Networks ……………………………………….. 124 Voice Communications over WLAN on Campus: A Survey ………………………………………………………….. 130 A 3-Dimensional Object Recognition Method Using Relationship of Distances and Angles in Feature Points ………………………………………………………………………………………………………………………… 137 Disparity Map Estimation using Local Gabor Wavelet under Radiometric Changes ……………………. 148 The Effect of Information Systems in the Information Security in Medical Organization of Shaharekord ………………………………………………………………………………………………………………………….. 157 Provide a Hybrid Approach to Manage Packets Motion in Order to Congestion Control in MANET ………………………………………………………………………………………………………………………. 167 Improvement of Load Distribution and Prevention of Congestion in Ad-Hoc Wireless Networks Using Classification and Swarm Intelligence ………………………………………………………………. 173 Platforms for Use Integrated Resources Formative Processes in E-learning …………………………….…. 181 Learning Management Systems: A Comparative Analysis of Open-Source and Proprietary Platforms ……………………………………………………………………………………………………………………………….…… 187 A Comparative Assessment of e-Learning Platforms ………………………………………………………………….. 193 EDA as a Discriminate Feature in Computation of Mental Stress ………………………………………………… 199 New Approach to Simulate Operation Rooms based on Medical Devices used in Surgical Procedures ……………………………………………………………………………………………………………………. 205

Proceedings of Second International Conference on Digital Information Processing, Data Mining, and Wireless Communications (DIPDMWC2015), Dubai, UAE, 2015

Increasing the Target Prediction Accuracy of MicroRNA Based on Combination of Prediction Algorithms Mohammed Q. Shatnawi, Jordan University of Science and Technology Faculty of Computer and Information Systems Computer and Information Systems Irbid, Jordan [email protected]

ABSTRACT MicroRNA is an oligonucleotide that plays a role in the pathogenesis of several diseases (mentioning Cancer). It is a non-coding RNA that is involved in the control of gene expression through the binding and inhibition of mRNA. In this study, three algorithms were implemented in WEKA software using two testing modes to analyze five datasets of miRNA families. The data mining techniques are used to compare the interactions of miRNA-mRNA that it either belongs to the same gene-family or to different families, and to establish a biological scheme that explains how the biological parameters are involved or less involved in miRNA-mRNA prediction. The factors that were involved in the prediction process includs match, mismatch, bulge, loop, and score to represent the binding characteristics, while the position, 3’UTR length, and chromosomal location and chromosomal categorizations represent the characteristics of the target mRNA. These attributes can provide an empirical guidance for study of specific miRNA family to scan the whole human genome for novel targets. This research provides promising results that can be utilized for current and future research in this field.

KEYWORDS miRNA, chromosome, prediction, genome, disease, biology

1 INTRODUCTION The cell, the basic unit of a living organism, has an extraordinary ability to reproduce, grow, respond to

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Mohammad S. Alhammouri, Kholoud Mukdadi Jordan University of Science and Technology Computer Science Department Irbid, Jordan [email protected] , [email protected]

stimuli, and exchange a wide range of materials (metabolites) with its surrounding environment. All these tasks have been found to be heritable in nature from parent cells to progenies, and have been explained through a molecular model that has become the dogma of molecular biology. The majority of the functions inside the cells are carried out by enzymes (Proteins and sometimes RNA). These components are orchestrated and controlled by genes. In simple terms, genes are heritable codes in the form of DNA sequence that can be transcribed into mRNA and then translated into proteins. This process is described as gene expression. Gene expression determines the fate of a cell because it controls the types of RNA and Proteins and also their exact amounts in a certain cell. Thus, all the cells in a human being contain the same DNA (genes) but due to differences in gene expression, a cell could become either a heart or a brain, or a muscle cell (with few exceptions). Genes are complex information systems that are chemically represented through a double strandedchain of DNA. Each strand of DNA is composed of deoxyribose/phosphate backbone and a nitrogen base (either G: guanine, C: cytosine, A: adenine or T: thymine). The nitrogen bases - are bind to each other chemically through hydrogen bonds in pairs of G-C and T-A. The structure of DNA allows for the two strands to separate, thus it can replicate using the complementarily property without losing its information. The information is stored in DNA in the form of base sequence. The stored information in genes is transcribed (i.e. copied) by an enzyme into a single stranded chain of RNA, which it is composed of ribose/phosphate backbone

1


and a nitrogen that can be of different types such as, (either G:guanine, C:cytosine, A:adenine or U:uracil). The nitrogen bases bind to each other chemically through hydrogen bonds in pairs of G-C and U-A. RNA can be either carrying a coding sequence (i.e. mRNA which is translated into protein) or a non coding sequence (other types of RNA). Finally, Proteins is composed of amino acids, and are constructed by ribosome according to the genetic code in a process called translation. Each of the 20 amino acids is represented in the genetic code as three nitrogen bases. Proteins are the functional component of the cell; they play a role in building the structure of the cell, as enzymes (catalyzing chemical reactions), as signaling molecules (hormones), or transport vehicles, etc [1]. 1.1 Messenger RNA (mRNA) An mRNA transcript is composed of several components:  5' untranslated region (5'UTR), which is a sequence of variable length that plays a role in the binding between ribosome and mRNA at the beginning of translation process. This UTR region may also contain miRNA binding sites as well. Thus it could play a role in the regulation of gene expression.  AUG codon: Start of translation site. AUG is translated into the amino acid methionine.  Coding domain sequence (i.e. CD or CDS) is the part of mRNA that is translated into proteins.  Stop codon: any of the three codons UAG, UGA and UAA. These codons do not code for proteins and they signal the end of the translation process.

transcript. A longer poly A tail and 3'UTR will increase the stability of mRNA against degradation (i.e. increase the half life of mRNA) [1]. 1.2 Micro RNA ( miRNA ) miRNA can be defined as a class of short noncoding RNAs,that are approximately 21 nucleotides (nt) in length, “which was first found in 1993 and several studies carried out on it a decade ago”.?? These are post-transcriptional genetic regulators, which have a big role in eukaryotic gene expression. It is not easy to predict human miRNA target which exist in some mismatch, gaps, and G:U wobble pairs. In addition, the human miRNA have multiple targets on the same mRNA [2]. 1.3 Regulation of Gene Expression Gene expression is regulated through a wide range of factors along with the different steps of transcription and translation. The first step of regulation occurs at the chromatin level. Chromatin is the natural packaging of DNA with special proteins that protects and controls genes inside the cell through several chemical modifications of the chromatin Histon proteins (e.g. Histon acetylation activates genes while methylation represses them etc.) and DNA itself. DNA methylation causes gene silencing. Another level of control occurs at the DNA sequence level though the promoter and other upstream regulatory elements which act as a binding site ,that is (called cis-elements) for transcription factors (trans-elements) which are complexes of proteins that recruit or repress the enzymes of transcription (i.e. RNA polymerases).

 3' untranslated region (i.e. 3'UTR): a sequence of variable length that plays a role in the stability of the mRNA. 3'UTR also contains sites for binding of miRNA sequence. Thus it plays a role in regulating of gene expression.

After transcription, mRNA levels are controlled by its half-life through several factors. (miRNA) is believed to be a key player in the control of mRNA levels. miRNA binds to mRNA in a specific fashion and recruits a number of enzymes responsible for mRNA degradation.

 Poly A tail: a sequence of adenines that also plays a role in the stability of mRNA

Finally, which expression can be regulated at the protein level either by chemical modifications of the

ISBN: 978-1-941968-26-0 ©2015 SDIWC

2


proteins (eg. phosphorylation) or by a feedback loop such as in the case of transcription factors that control their own genes. [1] 2 FUNCTIONS OF MIRNA (miRNA) plays a great role in Arranging a large number of target genes in the gene expression. Most of human miRNA genes have been defined and combined to realize a stable number of functions. Let-7 and Lin4 miRNA are considered as some of the first discovered miRNA genes in the C.elegans worm. Talking about human and other vertebrate cell lines, tumor suppression, antiviral defense, adipocyte differentiation and susceptibility to cytotoxic T-cells include some miRNA genes [2]. The MicroRNAs play important and critical roles in genetic human diseases [3, 4]. Such as, breast cancer [5] and heart diseases [6]. Nevertheless, the mechanism of gene expression that is regulated by miRNA remains ambiguous. 3 PROBLEM STATEMENT Predicting miRNA-mRNA interaction by experiment is highly costly in terms of time, labor and money for biologists. The number of miRNA and mRNA studies is increasing on a daily basis. Attempts to design biological experiments is always rendered outdated once the experiment is about to start. In addition, experiments may lack the element of discovering the novel miRNA-mRNA interactions. Thereby, data that is stacking in biological databases should be invested to discover new miRNA-mRNA interactions. The problems in miRNA-mRNA target prediction can be summarized in the following points:  Previous studies in the field of miRNA– mRNA target prediction are scattered in different aspects of the binding process such as sequence complementarity, the binding energy (thermodynamics), etc. This limits the involvement of the factors that play a role in the prediction.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

 Previous studies often use one classification method, which gives great weight to specific features on the account of others.  miRNA are uniquely involved in complex cellular pathways that include well organized and controlled networks of genes. These networks may or may not “cross-talk” with each other. This fact was not taken into consideration in the previous studies and miRNAs were studied in bulk and not as separate families. 4 PROBLEM SOLUTION Several attempts were used to apply bioinformatics techniques in order to find an optimized algorithm that can be used to explain miRNA-mRNA interaction mathematically as well as biologically. At mean time, there are several studies attempting to optimize miRNA-mRNA prediction from different aspects ranging from RNA sequence to binding energy (thermodynamics). Such algorithms can be used to screen the genome for new mRNA targets and predict miRNA-mRNA relationships that can be of a great benefit to the treatment and diagnosis of many diseases. In this study, the following solutions are taken into consideration for the prediction of miRNA-mRNA relationships:  Factors from different aspects of miRNAmRNA binding are given in this study, such as sequence complementarity, the binding energy (thermodynamics), etc..  Three classification methods (i.e. decision tree, naïve Bayes and support vector machine) are used.  In order to shed the light on the uniqueness of miRNAs five families of miRNA;they are studied as one collection and each on as a standalone family. 5 KNOWLEDGE DISCOVERY IN DATABASE PROCESS Knowledge discovery in database (i.e. KDD) is a very important process, which is the general process

3


of converting raw data into useful information. The data mining is an integral part of this process [7]. Therefore, there are enormous and massive collection of data that is stored about the genes, proteins, and other vital information for each human being. As a result, KDD process can be applied to extract information, patterns, and new rules using different techniques [8]. 6 OBJECTIVES The main goal of this study is to use data mining for predicting the miRNA-mRNA interaction through the implementation of the following objectives:  Collecting the miRNA and mRNA data from databases. These data include biological parameters that are related to sequence, chromosomal location, structure folding and previous known interaction scores.

not yet understood. The establishing of new parameters that are involved in this interaction gives more light attention to the biology behind the process of gene regulation. 8 RELATED WORK In the literature, there are several researches that have been done on the miRNAs to predict their putative target mRNAs. They have been classified into different categories: researches based on computational method and probabilistic models, and researches based on machine learning methods. Authors in [9] introduces a probabilistic model to show the binding preferences of miRNA and its predicted target. This model transforms an aligned duplex to represent a new sequence and used a Variable Length Markov Chain (VLMC) to determine the possibility of this sequence.

 The use of different data mining techniques to study and compare the interaction of miRNA that is either belonging to the same genefamily or to different families.  Establish a biological scheme that can explain how the biological parameters are involved( or less involved) in miRNA-mRNA prediction. 7 SIGNIFICANT CONTRIBUTIONS This study provides an insight to the biological parameters that are involved or neglected in miRNA-mRNA target prediction, and shed some light on the mechanisms that are underlying gene silencing in cancer cellular pathways. This study is rather significant from a clinical perspective; the establishment of a good miRNAmRNA prediction tool can help in discovering novel gene interactions, which can open the gate for new drug targets, and novel mutated disease genes. Thus, pushing forward the process of disease treatment and diagnosis is in progress. On the other hand, this field is still in its infancy and the nature of miRNA-mRNA interaction is still

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Figure 1. A hierarchy of miRNA–mRNA duplex alphabet.

In [10], Chenghai Xue et al. Proposed a computational method to find the functional miRNA–mRNA regulatory modules (FMRMs) and to collect the miRNA in normal case and prostate cancer as a case to study the method contains groups of miRNAs and their putative target mRNAs under specific conditions. This computational method has successfully identified down-regulated patterns of mRNAs targets that are associated with prostate cancer and mRNAs associated with normal cases. Briefly, after preparing the dataset, authors applied association’s algorithms in data mining to

4


identify the biologically related miRNA–mRNA groups.

investigate the miRNA-mRNA relationship by weka 3.6 software as shown in Figure 3.

In [13], William Ritchie et al. proposed an approach for the determination of putative miRNA targets based on a comparison between expression data of miRNAs and that of mRNAs using luciferase reporter assay. The miRNAs can decrease the expression level of targeted genes with direct correlation or indirect correlation between them. The success of this model was limited because the expression scalability of miRNA and mRNA was large. In addition, there are indirect functional relationships between two molecules.

All miRNA types in this study had the same length (22 nt) and were previously shown to play role in causing cancer. For instance mir-21 is involved in breast cancer pathogenesis.In order to lower the false positives in this study, only the highest scored miRNA binding site (out of three) was included.

Xiaofeng Song et al. in [14] proposed a computational method that is called microDoR to identify the mechanism of gene silencing by miRNA in humans, after they analyzed many features to find which are correlated with gene inhibition by miRNA. In [15], Scott Younger et al. used computational methods for predicting possible miRNA targets through gene promoters and showed those promoters. Although, they are not conventionally linked to miRNA, they are strong candidates for miRNA regulation. In [16], Alain Sewer et al. used a computational method to develop a program to approximate the pre-miRNA content and to predict the site of precursor miRNAs in genomic sequences. 9 RESEARCH METHODOLOGY 9.1 Overview Most of the previous miRNA studies focused on the target prediction of miRNA binding using many features such as data mining or statistical techniques; trying to help and guide the experiments in the laboratory. Therefore, this study focuses on finding a correlation between the miRNA target sites of specific types of miRNA, namely; let-7a, let-7b, let-7c? family, mir-21 and mir-122 and the chromosomal location [i.e. q: long arm of chromosome and p : short arm of chromosome], the nucleotide sequence, binding and thermodynamic features of miRNA and mRNA. The dataset was collected from miRNA target prediction database (i.e. database using miRANDA algorithm) and then three techniques of data mining are used to

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Figure 2. The research methodology.

9.2 Data Set The result of this study depends on the quality of the datasets. Therefore, the miRNA features are collected from miRTarBase database that contain more than 3500 MTIs (miRNA–target interactions). The database content is updated manually by surveying pertinent literature using filter research articles related to functional studies of miRNAs. In general, reporter assays, western blot, or microarray experiments with over expression or knockdown of miRNAs validate experimentally the gathered MITLs. 4,270 miRTarBase currently contains experimentally verified MTIs; it contains between 669 miRNAs and 2,533 target genes amongst 14

5


species. The miRTarBase provides the last updated collection compared with other similar miRNA databases. In addition, it contains the largest amount of validated MTIs [18].

for each miRNA representing one of the categories to be as a class label with the above mentioned attributes. 9.3 Evaluation of the Results

9.3 Classification Techniques This study uses the following classification technique:  Support Vector Machine (SVM): supervised learning machine tool that is used to classify a sample of data set into two predefined classes, based on statistical analysis [30].  Naive Bayes Classifier: a simple supervised learning machine tool that employs Bayes’ theorem with independence assumptions among features [31].  Decision Tree Learning: supervised machine learning tool that is used as a predictive model to represent all effective (i.e. higher weight) decisions. Tree Leaves represent the possible classes while the edges represent conjunctions of features [32]. The classification is performed in two steps.  Step 1: Classification Using Target Gene Chromosome Location as a Class Label. The data mining classification algorithms are applied for each kind of miRNA where the position attribute is the class label. However, the results are reported accuracies below 50%. Therefore, we made the mRNA target gene Chromosome Location attribute to be the class label, which started to provide improvement in the accuracy. Three algorithms of classification from the weka 3.6 software are used (decision tree –J48, naïve Bayes and support vector machine -SMO) where each algorithm is applied twice on each file with different ways of data training split: Crossvalidation and percentage split, in which they are the default settings.  Step 2: Classification after Addition of Chromosome Categories to the Dataset. After that, a new feature is selected to add to the features set. The mRNA target genes Chromosome Number are grouped to four categories. Therefore, four files were made

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Accuracy was used for evaluating the experiments. Accuracy was represented by the percentage of the correctly classified identified records according to the following equation: Accuracy (Acc) = (TP + TN)  (TP + TN + FP + FN) Where; TP: number of true positives, TN: number of true negatives, FP: number of false positives and FN: number of false negatives [7]. 9.4 Experiment and Results In this study, three algorithms in WEKA software are used (e.g. decision tree –J48, naïve Bayes and support vector machine -SMO) with default setting of data training split way, Cross-validation 10 folds or percentage split 66%. In the first step, the aforementioned three algorithms were performed on all miRNA family datasets using the position attribute as a class label. Unfortunately, the accuracy of the results is lower than 50%. However, after changing the class label to the mRNA chromosome location attribute, the result has improved. Using the dataset , we applied classification methods using miRNA position attribute with chromosome location attribute as a class label. The highest accuracy that is reported using decision tree algorithm in Step 1 was for miR-21 (Acc=73.68%) and let-7a (Acc= 69.23%) in 66% test mode. Whereas, in the 10 fold test mode the highest accuracy was reported in miR-21 (Acc=65.45%) and miR-122 (Acc=60.47%) as shown in Figure 3A. A clear difference between 66% and 10 fold test modes was only seen in let-7a by a shift of 20%. The highest accuracy that is reported using naïve Bayes in Step 1 was seen in miR-122 (Acc=60%) and miR-21 (Acc=57.89%) in 66% test mode. Whereas, in the 10 fold test mode the highest accuracy was seen in miR-21 (Acc=67.27%) and let-7 family (Acc=57.21%) as

6


shown in Figure 3B. Using the support vector machine algorithm almost, all miRNA families provide equal accuracies that are over 50% except for let-7 family in 66% testing mode. Whereas, in 10 fold test mode all accuracy values were equal and over 50% except the let-7a as shown in Figure 3C.

experimentally. In this study, data mining techniques were used to classify a number of characteristics involved in miRNA binding and the mRNA targets themselves. Five families of miRNAs that are involved in cancer pathways have been analyzed in this study. The results can be summarized as follows: The use of decision tree in miRNA-mRNA target prediction shows that each miRNA family behaves in a unique way when it comes to binding features with or without chromosomal categorization: 

Figure 3. The accuracy of miRNA-mRNA predictions in step 1 (before the addition of chromosome categorization) according to (A) Classification using decision tree. (B) Classification using naïve base. (C) Classification using support vector machine.

10 CONCLUSION AND FUTURE WORK miRNA research has been developed progressively in the past few years. Prediction of miRNA-mRNA target was attempted both computationally and

ISBN: 978-1-941968-26-0 ©2015 SDIWC

The highest accuracy reported without chromosomal categorization was for miR-21 and let-7a in 66% test mode and for miR-21 and miR-122 in the 10 fold test mode and with chromosomal categorization the highest accuracy reported was seen in miR-21 and let7a in both testing modes.  The decision tree of let-7a showed the greatest weight to the mismatch followed by position and score without chromosomal categorization. While in step 2 the class categorization became the root for the tree followed by the matches and bulges.  The decision tree of miR-21 was the same with and without chromosomal categorization. The root was the match attribute followed by 3’UTR length.  The decision tree of miR-122 shows a clear complication and branching when the class categorization was added. The tree develops from one weight attribute in step 1 which was the 3’UTR length into a more complicated branching in step 2 including many attributes.  Binding features such as the match, mismatch, and bulge as well as the length of the 3’UTR was shown to play major role in the classification of targets. In addition, the chromosomal source of the target that is represented here by the class categorization contributed in the accuracy of the test. The use of naïve Bayes without chromosomal categorizations showed the highest accuracy in miR-21 and miR-122 families in 66% test mode and for miR-21 and let-7a in the 10 fold test mode. Whereas, when the chromosomal class categorization was used, the highest accuracy was

7


seen in “All miRNA” and miR-21 in 66% test mode and in miR-21 and let-7 family in the 10 fold mode.

[4]

When using support vector machine without chromosomal categorization almost all miRNA families showed similar accuracies (over 50%) except for let-7 family in 66% testing mode and let7a in 10 fold testing mode. When chromosomal categorization was used, the highest accuracy reported was seen in let-7a and miR-21 in 66% test mode and let-7a and miR-21 in 10 fold test mode.

[5]

Out of 26 features included in this study, only 9 features were retained. The rest of the features were eliminated either due to the low number of miRNAs included in this study or because they did not have any effect on the experimental results. The factors that were involved in prediction including match, mismatch, bulge, loop, and score represent the binding characteristics, while the position, 3’UTR length, and chromosomal location and chromosomal categorizations represent the characteristics of the target mRNA. In the future, several attributes such as the match, mismatch, bulge, and 3’UTR length can provide a threshold-based empirical guidance for study of specific miRNAs to scan the whole human genome for novel targets. In addition, naïve Bayes and support vector machine can be used to test the new attributes, especially the ones involved in the source of the target mRNA (i.e. chromosomal based attributes). New findings in the field of miRNA have the potential to revolutionize the study of many diseases. Many of the known miRNA are under focus now for targeted medicine and research is now ongoing in the field of using these miRNA as drugs for treatment of different types of cancer and diagnosis.

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16]

[17]

[18]

REFERENCES James Watson, Tania Baker, Stephen Bell, Alexander Gann, Michael Levine, Richard Losick. (2003) Molecular Biology of the Gene, Fifth Edition, Pearson (Benjamin Cummings) Publishing. [2] Mohammed Abba, Heike Allgayer. MicroRNAs as regulatory molecules in cancer: a focus on models defining miRNA functions. Drug Discovery Today: Disease Models 2009; 6(1): 13-19 [3] Jürgen Wittmann, Hans-Martin Jäck. Serum microRNAs as powerful cancer biomarkers 2010; 1806: 200–207. [1]

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Erik A.C. Wiemer. The role of microRNAs in cancer: No small matter. Euro pean journal of cancer 2007; 43: 1529 –1544. Cathy A. Andorfer, Brian M. Necela, E. Aubrey Thompson and Edith A. Perez. MicroRNA signatures: clinical biomarkers for the diagnosis and treatment of breast cancer. Trends in Molecular Medicine 2011; 17(6): 313-319. Haverich, Carina Gross, Stefan Engelhardt, Georg Ertl, Johann Bauersachs van Laake, Pieter A. Doevendans,el at. MicroRNAs in the Human Heart : A Clue to Fetal Gene Reprogramming in Heart Failure. American Heart Association 2007; 116: 258-267. Usama Fayyad, Gregory Piatetsky-Shapiro, and Padhraic Smyth. From Data Mining to Knowledge Discovery in Databases. AI MAGAZINE 1996; 17(3): 37-54. Pang-Ning Tan,Michael Steinbach and Vipin Kumar. Introduction to Data Mining. International Edition. Boston 2006; ISBN-10: 0321420527 Hasan Og˘ul, Sinan U. Umu, Y. Yener Tuncel and Mahinur S. Akkaya. A probabilistic approach to microRNA-target binding. Biochemical and Biophysical Research Communications 2011; 413: 111-115. Bing Liu, Jiuyong Li and Anna Tsykin. Discovery of functional miRNA–mRNA regulatory modules with computational methods. Journal of Biomedical Informatics 2009; 42: 685–691. Wan Hsieh, Hsiuying Wang. Human microRNA target identification by RRSM. Journal of Theoretical Biology 2011; 286: 79–84. National Chiao-Tung University [Internet]. Taiwan: National Chiao-Tung University; Available from: http://www.stat.nctu.edu.tw/_hwang/website_wang%20ne w.htm William Ritchie, Megha Rajasekhar, Stephane Flamant, John Rasko. Conserved Expression Patterns Predict microRNA Targets. PLoS Computational Biology 2009; 5(9). Xiaofeng Song , LeiCheng , TaoZhou , XuejiangGuo , XiaobaiZhang , el at. Predicting miRNA-mediated gene silencing mode based on miRNA-target duplex features. Computers in Biology and Medicine 2011; 42(1): 1-7. Scott T. Younger , Alexander Pertsemlidis , David R. Corey. Predicting potential miRNA target sites within gene promoters. Bioorganic & Medicinal Chemistry Letters 2009; 19: 3791–3794. Alain Sewer, Nicodème Paul, Pablo Landgraf, Alexei Aravin, el at. Identification of clustered microRNAs using an ab initio prediction method. BMC Bioinformatics 2005, 6:267. Xingqi Yan, Tengfei Chao, Kang Tu, Yu Zhang, Lu Xie, Yanhua Gong, Jiangang Yuana, Boqin Qiang and Xiaozhong Peng. Improving the prediction of human microRNA target genes by using ensemble algorithm. FEBS Letters 2007; 581: 1587–1593. Sheng-Da Hsu, Feng-Mao Lin, Wei-Yun Wu, Chao Liang, Wei-Chih Huang, Wen-Ling Chan, et al. miRTarBase: a database curates experimentally validated microRNA–target interactions. Nucleic Acids Research 2011;39: 163-169

8


Clinical Prediction of Teeth Periapical Lesion based on Machine Learning Techniques An Experimental Study Yasmine Eid Mahmoud1, Soha Safwat Labib2 and Hoda M. O. Mokhtar3 1,2 Faculty of Computer Science, October University for Modern Sciences and Arts, Giza, Egypt 3 Faculty of Computers and Information, Cairo University, Giza, Egypt 1 [email protected], [email protected], [email protected] ABSTRACT Dentists used to diagnose teeth periapical lesion according to patient’s dental x-ray. But most of the time there were a problematic issue to reach a definitive diagnosis. It takes too much time, case and chief complaint history needed, many tests and tools are needed and sometimes taking too many radiographs is required. Even though, sometimes reaching definitive diagnosis before starting the treatment is difficult. Therefore, the objective of this research is to predict whether the patient has teeth periapical lesion or not and its type using machine learning techniques. The proposed system consists of four main steps: Data collection, image preprocessing using median and average filters for removing noise and Histogram equalization for image enhancement, feature extraction using segmentation and expectation maximization algorithm, and finally machine learning (classification) using Feed Forward Neural Networks and K-Nearest Neighbor Classifier. It has been concluded from the results that the K-Nearest Neighbor Classifier performs better than Feed Forward Neural Network on our real database.

KEYWORDS Image Segmentation; Expectation Maximization; Histogram Equalization; Classification

1 INTRODUCTION No one can deny the importance of information technology (IT) in our lives as it makes our lives easier each day. Nevertheless, the use of IT techniques has an important role in many domains. In the dental domain, IT can help the dentists to predict and treat the infection the right way. Dental caries (tooth decay) is considered one of the most widely spread infections in human

ISBN: 978-1-941968-19-2 ©2015 SDIWC

kind. If any caries lesion is left untreated, bacteria will be no longer confined to hard tooth structure, it will go down to the soft tissue underneath, which is the dental pulp, causing pulp infection as shown in Figure 1 [15]. In dental world there are many types of Periapical abscess like Acute Periapical Abscess, Acute Apical Periodontitis, Chronic Periapical Abscess, periapical granuloma and etc [1].

Figure 1: Periapical abscess and cellulitis

Acute Periapical Abscess (APA), also known as acute apical abscess, acute dentoalveolar abscess or acute periradicular abscess, is a highly symptomatic painful inflammatory response of the periapical connective tissues. It is caused by pulpal necrosis which originates from pulpal tissues which initiate an inflammatory response to trauma or caries and may subsequently lead to pulpal necrosis [1]. Radiographically, the appearance of the periodontal ligament space ranges from within normal limits, to slightly thickened, to large periapical radiolucency [2].

9


The initial periapical response to bacterial presence within the canal or to bacterial invasion of the periapical region will be an acute inflammatory response, known as primary acute apical periodontitis (AAP) [3]. Chronic Periapical Abscesses (CPA) is an inflammatory reaction to pulpal infection and necrosis characterized by gradual onset, little or no discomfort and an intermittent discharge of pus through an associated sinus tract [3]. Teeth Periapical lesion needs special medication where surgical treatment might be needed according to the infection type. Today, dentists diagnose the periapical lesion according to patient’s dental x-ray. Most of the time, dentist misdiagnose the type of periapical lesion because it is difficult for human naked eyes to easily observe and classify the type. Hence, the true type is usually discovered through surgical treatment. Misdiagnosing or even late diagnosis may result in the spread of the infection. Image processing and machine learning are thus needed for solving such problem. Proper use of technology and techniques can speed the diagnosis process and help to achieve more accurate diagnosis. 2 PRELIMINARIES 2.1 Average Filter To minimize the frequency of noise in images average or mean filter is needed. The extent of intensity variation is minimized among neighboring pixels through this filter [5]. The algorithm works on image matrix by calculating the average value of the pixel neighbors including itself and the resulting value will be replaced with this pixel. Average filter is to some extent similar to the convolution filter [5], where the convolution filter also relies on calculating the mean for kernel to represent the shape and size of the vicinity to be sampled. Frequently the main filter practice is to use 3x3 square kernel, but if high smoothing is needed, then larger kernels should be used like 5x5 or 7x7 etc. The average is calculated by adding all

ISBN: 978-1-941968-19-2 ©2015 SDIWC

the values of the surrounding pixels and taking the average of them. Then, the central pixel value will be replaced with the resulting average. The algorithm takes the input image and outputs it after smoothing or filtering [5]. 2.2 Median Filter For improving the results of the later processing median filter is needed to reduce image noise. It preserves edges even after removing noise so it’s a nonlinear digital filtering technique. There is one dimensional median filter that works on signals by running through it entry by entry. Each entry will be replaced by the median of neighboring entries. Beside the one dimensional there is twodimensional median filter that works on images by using a square window; for each window the central pixel will be replaced by the median value of its surrounded neighbors. The median is calculated by sorting the given window, then if the length of the window is odd so simply the median will be the middle value. Otherwise, if the length is even then there is more than one possible median [6]. 2.3 Histogram Equalization Distributing the gray level within an image is the goal of histogram equalization in order to equally likely occurring pixels’ gray level. It will decrease the contrast images and enhance the brightness and contrast of dark. The main job of histogram equalization is to increase contrast. Some times in some cases histogram result worse and in this cases the contrast decreased [8]. Improving the quality of the image is a benefit reached by using histogram modification. It will gather near lower gray level if it is applied on dark image whereas if it is applied on brighter image, it will gather near higher gray level. Therefore, histogram will not spread equally on low contrast image and it will spread equally among gray level on high quality image. The algorithm consists of four main steps. First, perform running sum of histogram values i.e. CFD (Cumulative frequency

10


distribution), then divide each result with total no. of pixels, then multiply the result by maximum gray level Value, and finally use One-to-One correspondence, map gray level Values in result obtained from multiplication (third step) [9]. 2.4 Image Segmentation The foundation of object recognition and computer vision is the concept of image segmentation. It is the procedure of subdividing a digital image into multiple regions that consists of set of pixels that have the same properties or characteristics which are given different labels for representing different regions. Simplifying the representation of an image into something that is more meaningful and making image analysis simpler is the main goal of segmentation. Setting objects and boundaries in images is done using image segmentation. The segmentation types are soft segmentation and hard segmentation. Allowing overlapping in regions or classes called soft segmentation whereas forcing a decision of whether a pixel is inside or outside the object called hard segmentation [7]. 2.5 Expectation Maximization Computing the maximum-likelihood estimates when the observations are incomplete is the EM (expectation-maximization) algorithm. Calculating the density estimation of data points in an unsupervised setting is the use of EM. It simplifies difficult maximum likelihood problems. It has two main steps, Estep used for computing the expectation then Mstep is used to maximize the last step and EM iteration steps continue until convergences occur. The research of the parameter achieving the maximum likelihood is the concept of EM algorithm. The stopping conditions of the algorithm are reaching the maximum number of iterations to limit the time of calculation or a lower mistake happens. Because it reposes on the calculation of the complete data, it is easily put in application. The problem of EM is, it doesn’t directly fit in spatial modeling and

ISBN: 978-1-941968-19-2 ©2015 SDIWC

therefore it can be sensitive to noise and intensity in identities [10]. Assume that the random observation is the intensity of the nth pixel, and all the pixels of the ith image cluster have a mean intensity and variance . The K mixed Gaussian distribution can be written as [11] ( ) ( |) (1) ( |)

∑

√

{

(

)

}

Using the maximum likelihood estimator (MLE), the parameters , and may be estimated. Desirable asymptotic properties are known to be had by the MLEs. The MLEs are used for estimating parameters in mixture models. The loge of the likelihood function is given by [11] ∑ ( | )} (2) ( ) ∑ {∑ Where the total number of pixels considered is denoted as N. Equating the first partial derivatives of (2) with respect to unknown parameters to zero obtain the ML equations. Solving for parameters in the likelihood equations obtain the ML estimates. The estimation procedure becomes easier by going through the EM algorithm. Let ( | ) be the subsequent probability that the random observation belongs to the ith class. In the algorithm, this subsequent probability is updated by the E-step given the latest estimates of ̂ ̂ ̂ ; i.e. [11] ̂( | )

̂ ̂( | ) ∑ ̂ ̂( | )

(3)

The likelihood equations of (2) contain the subsequent probability ̂( | ); therefore, by substituting (3) into these equations, the M-step simply computes the unknown parameters as follow: [11] ∑ ̂( | )

̂ ̂

̂

(4)

∑ ̂( | )

(5)

11


̂

̂

∑ ̂( | ) (

̂)

(6)

Until the sequence of parameter estimates becomes stable, the E-steps and M-steps are iterated [11]. 2.6 Feed Forward Neural Network Feed Forward neural network is an Artificial Neural Network. In Artificial Neural Network, The problem information is distributed in neurons and these neurons are connected by weights of links. For creating the desired mapping, the neural network has to be prepared to regulate the connection weights and biases in society. For complex pattern recognition and classification tasks, the ANNs are mainly helpful. Learning from examples leads to the ability of reproducing arbitrary non-linear functions of input, and it is suitable for pattern classification problems because of the highly parallel and regular structure of ANNs. The algorithm main goal is to minimize error by adjusting the weights of the network connections. By comparing the obtained outputs with the expected outputs of known inputs, the error is calculated. The weights are adjusted after the error is backward propagated until the first layer. The weights are continually adjusted since this process is occurred over and over. The training set is the set of data which enables the training. Until reaching an acceptable error or the maximum number of iterations the same set of data is processed many times during the training of a network [12]. 2.7 K-Nearest Neighbor Classifier The non-parametric algorithm used for classification is KNN. Previous knowledge about the structure of the data in training set isn’t needed in KNN classifier. Retraining isn’t needed if a new training pattern is added to existing training set. Its output can be taken as a subsequent probability of the input pattern belonging to a particular class. The confidence in prediction is improved as the k is increased [9].

ISBN: 978-1-941968-19-2 ©2015 SDIWC

KNN consists of training and testing phases, where in the training the data should be labeled while in the testing phase the data is unlabeled and the algorithm should identify the list of k nearest data points and classify their class. For classifying the class a majority vote needed to determine the class label using class labels of nearest neighbors. The algorithm main step is to calculate the distance between the stored records and the unknown one. Distance is calculated using Euclidean distance (Equation 7) [13]. ( ) √∑ ( ) (7) 3 DATA DISCRIPTION In this work we used real data obtained for the dental hospital at MSA University – Egypt. Our data consists of 201 dental x-ray images for teeth periapical lesion. The data was used in two ways. The first way, the data is divided into four classes: class 1 APA (Acute Periapical Abscess), class 2 AAP (Acute Apical Periodontitis), class 3 CPA (Chronic Periapical Abscess), and class 4 Normal. In other way the data is divided into two classes: class 1 infected and class 2 Normal teeth. 4 PROPOSED SYSTEM The architecture of the proposed approach consists of five main building blocks as shown in Figure 2. The first level is concerned with data collection. Level 2 focuses on image processing. Feature extraction is the core of level 3. Data is then classified in level 4. Finally, the approach evaluation is conducted in level 5.

12


Data Collection

Image Preprocessing 1. Remove Noise 2. Histogram Equalization

Feature extraction 1. Segmentation 2. Expectation Maximization

Classification

Evaluation Figure 2: Proposed System Diagram

4.1 Data Collection The data are collected from dental clinics of Modern Science and Arts University. This collected data were in the form of hard copy xray, these x-rays were photo copied using camera and x-ray viewer. The used camera was a digital camera with features 7.2 Mega-Pixel CCD captures enough detail for photo-quality 15 x 20-inch prints, produce images 3072×2304 with file format JPEG. 4.2 Image Preprocessing There are a number of factors affecting the teeth detection. These factors include: capturing quality, capturing resolution, x-ray quality and noise affected x-ray. The importance of the preprocessing stage lies in its ability to treat some of the problems that may occur due to some of the factors presented before. Enhancing the x-ray image to be prepared for the next stage may be resulted using the preprocessing techniques. It is better to have an effective preprocessing step in order to achieve higher detection rates. [4] In this paper, image preprocessing is achieved in two phases. The first phase was

ISBN: 978-1-941968-19-2 ©2015 SDIWC

removing noise where spatial filters were used. Those filters like average and Median filter. Then, the second phase was enhancing the filtered image using histogram equalization. 4.3 Feature Extraction After image preprocessing now we need feature vectors that describes our data. Therefore, image should be converted to meaningful feature. We firstly used image segmentation that converts the image into set of segments and each segment contains pixels that have the same properties or characteristics. We applied it in four different ways: 4, 8, 12, and 16 segments for each image. Then after segmentation we used expectation maximization for substituting each segment with one value. In our testing, we used the average (mean) value resulted by the expectation maximization because it is the best representative for the segment pixels. 4.4 Classification Supervised learning is needed in this work. The system is expected to find out common properties of the different classes, and what differentiates them, in order to make correct classification for other new unseen cases. This type of supervised learning is known as “Classification”. The used classification algorithms are Feed Forward Neural Network and K- Nearest Neighbor Classifier. 4.5 Evaluation Evaluation is the important that a diagnostic test is correctly performed. Using two classes’ data set, we used to calculate the accuracy as follows [14]: ( (

) )

(8)

Using four classes’ data set, we used to calculate the accuracy as follows [14]: (9)

13


5.1 Two-classes Data set As shown in table 1 we applied KNN and Feed Forward Neural Network algorithms on the four ways resulted from feature extraction. After finishing classification we calculate the accuracy for each way. In KNN classifier needs K-value, and we tested different values, the best results were the k=4 with Feature way of 4 segments and k=7 with feature way of 12 segments. In Feed Forward Neural Network classifier needs two parameters, the first is the number of layers and the number of neurons in each layer and the second is the number of iterations, and the best results were the layers are three with neurons 7 5 1 respectively and number of iterations is 100; we can conclude that the accuracy of using Feature way of 4 segments are better than other ways. From the results also, we conclude that the accuracy of applying KNN classifier using Feature ways of 4 or 12 segments are better than Feed Forward Neural Network. TABLE I.

RESULTS OF ACCURACY USING KNN CLASSIFIER AND FEED FORWARD

No. of Segments Classifier

4

8

12

KNN

95.65% K=4

82.61% K=5

95.65% K=7

91.30% K=6

86.96%

69.57%

73.91%

78.26%

Feed Forward NN

Results of Accuracy using KNN Classifier and Feed Forward

100% Acurracy

5 EXPERIMENTAL RESULTS This section is divided into 2 parts; discussing the results when applying the classification algorithms on A) Two-classes data set which differentiate whether infected or not. B) Four-classes data set which explain the infected periapical lesion type. The collected data will be divided into 88% of them will be training and 12% will be testing data.

80% KNN

60% 40% 0% 4

8 12 16 No. of Segments Figure 3

5.2 Four-classes Data Set As shown in table 2 we applied KNN and Feed Forward Neural Network algorithms on the four ways resulted from feature extraction. After finishing classification we calculate the accuracy for each way. In KNN classifier needs K-value, and we tested different values, the best results were the k=4 with feature way of 12 segments. In Feed Forward Neural Network classifier needs two parameters, the first is the number of layers and the number of neurons in each layer and the second is the number of iterations, and the best results were the layers are three with neurons 15 10 1 respectively and number of iterations is 100; we can conclude that the accuracy is not that good and using Feature way of 4 or 12 segments are better than other ways. From the results also, we conclude that the accuracy of applying KNN classifier using Feature ways of 12 segments is better than Feed Forward Neural Network.

16 TABLE II.

RESULTS OF ACCURACY USING KNN CLASSIFIER AND FEED FORWARD

No. of Segments Classifier

KNN Feed Forward NN

ISBN: 978-1-941968-19-2 ©2015 SDIWC

Feed Forwar d NN

20%

4

8

12

16

39.13% K=4

56.52% K=5

78.26% K=4

47.83% K=6

43.48%

30.43%

43.48%

17.39%

14


[6]

Results of Accuracy using KNN Classifier and Feed Forward

[7]

100%

Acurracy

80%

[8]

60% KNN

40% 20% 0% 4

8 12 16 No. of Segments

Feed Forwar d NN

[9]

[10]

[11]

Figure 4 [12]

6 CONCLUSIONS AND FUTURE WORK In this study, a new application is developed to help dentist diagnose type of teeth periapical lesion. The application system has four main steps: Image Preprocessing using median and average filters for removing noise and Histogram equalization for image enhancement, Feature extraction using segmentation and expectation maximization algorithm, and finally Machine learning (Classification) using K-Nearest Neighbor Classifier and Feed Forward Neural Network. The K-Nearest Neighbor Classifier is the better than Feed Forward Neural Network according to the results. Future work will concentrate on enhancing the results using optimization technique.

[13]

[14] [15]

Filter', International Journal of Computer Applications, pp. 25-28, 2014. P. Patidar, M. Gupta, S. Srivastava and A. Nagawat, 'Image Denoising by Various Filters for Different Noise', International Journal of Computer Applications, vol. 9, no. 4, pp. 45-50, 2010. P. Shivhare and V. Gupta, 'Review of Image Segmentation Techniques Including Pre & Post Processing Operations', International Journal of Engineering and Advanced Technology, vol. 4, no. 3, pp. 153-157, 2015. S. B. Rana and S. B. Rana, 'A Review of Medical Image Enhancement Techniques for Image Processing', International Journal of Current Engineering and Technology, vol. 5, no. 2, pp. 1282-1286, 2015. M. Kaur Khehra and M. Devgun, 'Survey on Image Enhancement Techniques for Digital Images', Scholars Journal of Engineering and Technology, vol. 3, no. 2, pp. 202-206, 2015. R. Ravindraiah and K. Tejaswini, 'IVUS Image Segmentation By Using Expectation-Maximization Approach', International Journal of Advanced Research in Computer and Communication Engineering, vol. 3, no. 2, pp. 5662-5664, 2014. J. Wu, J. Chen, X. Zhang and J. Chen, 'The Segmentation of Brain MR Images using Reformative Expectation-Maximization Algorithm', Int. J. Image Grap., vol. 10, no. 02, pp. 289-297, 2010. S. Vyas and D. Upadhyay, 'Classification Of Iris Plant Using Feedforward Neural Network', International Refereed Journal of Engineering and Science, vol. 3, no. 12, pp. 65-69, 2014. N. V. Chavan, B. D. Jadhav and P. M. Patil, 'Detection and Classification of Brain Tumors', International Journal of Computer Application, vol. 112, no. 8, pp. 48-53, 2015. J. Han and M. Kamber, Data mining. Amsterdam: Elsevier, 2006. D. Martin, 'Common Dental Infections in the Primary Care Setting - American Family Physician', Aafp.org, 2015. [Online]. Available: http://www.aafp.org/afp/2008/0315/p797.html. [Accessed: 28- Oct- 2015].

REFERENCES [1]

[2]

[3] [4] [5]

L. Walsh, 'Serious complications of endodontic infections: Some caustionary tales', Australian Dental Journal, vol. 42, no. 3, pp. 156-159, 1997. D. Matthews, S. Sutherland and B. Basrani, 'Emergency Management of Acute Apical Abscesses in the Permanent Dentition: A Systematic Review of the Literature', Journal of the Canadian Dental Association, vol. 69, no. 10, p. 660, 2003. K. Hargreaves, S. Cohen and L. Berman, Cohen's pathways of the pulp. St. Louis, Mo.: Mosby Elsevier, 2011. Y. Alginahi, 'Preprocessing Techniques in Character Recognition', Character Recognition, 2010. H. Prateek Singh, A. Nigam, A. Kumar Gautam, A. Bhardwaj and N. Singh, 'Noise Reduction in Images using Enhanced Average

ISBN: 978-1-941968-19-2 ©2015 SDIWC

15


Extraction of causalities and rules involved in wear of machinery from lubricating oil analysis data Daisuke Ide∗1 , Atsushi Ruike∗2 , Masaomi Kimura∗3 { Division of Electrical Engineering and Computer Science, ∗3 Department of Information Science and Engineering,} Shibaura Institute of Technology, 3-5-7 Koto-ku Toyosu, Tokyo 135-8548, Japan ∗2 Tribotex, 45-7 Yamaguchi, Nagakusa-machi, Obu-shi, Aichi 474-0052, Japan {∗1 ma15012, ∗3 masaomi}@shibaura-it.ac.jp ∗1

ABSTRACT Recently, methods in order to diagnose wear conditions of the equipments have been established. Lublicating oil analysis is one of these methods. However, since relations between events in wear are complex, its diagnosis relies on judgment by experts at this moment. In order to solve this problem, a purpose of this study is to support its diagnosis by generating a automatic diagnosis model. In this paper, we proposed a method that generate the model in order to predict wear conditions of the equipments. First, the causalities in wear were extracted from the diagnosis reports which experts described considerations for wear conditions of the equipments using text mining. Second, the equipments which has similar features were classified using clustering and the rule of each cluster was extracted using decision tree from analysis data related to lubricating oil and equipments. Finally, the models were generated by combining the causalities and the rules. Although the results of evaluation indicated that automatic diagnosis is possible, it will be necessary to diagnose the more detailed wear conditions of machinery in the future tasks.

KEYWORDS Text mining, Clustering, Decision tree, Lubricating oil analysis

1

INTRODUCTION

Wear of lubricated parts in machines can cause machine failure and degrade their performance. In order to prevent the severe wear, comprehension of states of the parts in the machine is essentially important. The traditional way to inspect a wear condition is direct check

ISBN: 978-1-941968-26-0 ©2015 SDIWC

of the lubricated parts by dismantling the machines. However, it is not realistic, since it takes costs of dismantlement itself and of stoppage of the machines. Therefore, it is important to establish a method to diagnose them without disassembly. Lubricating oil is used to decrease friction on lubricated parts of machines. The lubricating oil and wear particles have strong relationship. Since wear particles appear during the oil is used, the particles in lubricating oil suggests wear condition of the parts. A well-known method, “lubricating oil analysis”, uses this fact. It can specify wear location by discriminating a metal type of wear particles and can estimate wear of machinery by quantity and size of wear particles. In addition, it is possible to diagnose wear condition of mechinery and its cause by specifying the form of particles, which vary by various factors, e.g., the lubricating oil viscosity, acid number, additives and other surrounding enviroment. Since events (phenomenons) on the lubrication parts in the wear are numerous and their causality is extremely complicated, the lubricating oil analysis requires judgement of wear condition in consideration of this complex causality[1]. Since lubricating oil analysis is currently judged only by experienced experts, it has the problem such as lack of diagnosis time and the difference of results depending on experts. In order to solve it, it is desirable to realize the automation of equipment diagnosis. In this study, in order to obtain information useful for such automation, we applied text mining and data mining to extract causality and rules related to wear of machinery from lubricating oil analysis data. The data consisted of “analysis

16


data” which was measured for analysis fields related to lubricating oil and equipments, and “diagnostic report” which described consideration by experts for the equipments based on analysis data[2]. The objective of this study is to propose a method to generate prediction model of wear condition of machinery by extracting the causality and rules related to wear using historical lubricating oil analysis. Omori et.al proposed a method to extract a sentence including a description of the causality related to product defects[3]. Although discrimination of cause and result is important in causality, their method gives only the extracted sentence but does not identify what is cause (or result). In this study, for diagnostic report, we focused not on extraction of the sentence that includes a description of the causality but on identification of causes and results of causality therein [4]. For analysis data, the equipments with similar characteristics were classfied into a dataset, from which the rules were extracted by data mining[5]. We employed two-step clustering method to get the datasets, and used C&R Tree to obtain conditions that generate severe wear as regularity. Based on the results, we generated a model to diagnose a wear condition of machinery by combining the extracted causality with the rules. 2 2.1

of expressions of frequently appearing nouns or typographical error. In order to remove these nouns, we extracted the nouns whose frequency is more than 10% of the number of words in all diagnostic reports. In the second step, we identified words modified by nouns in the noun list. We extracted the words, if the noun means “the materialsubstance” and the modified word means its change or detection, if the noun means “the parts” and the modified word means its state, or if the noun means “the state” and the modified word means its extent or change as events. Table 1 shows an example of extracted events. In the third step, we extracted the description features of causality in diagnostic reports. When we confirmed sentences that contain more than two extracted events, we found that, in diagnostic reports, there were the frequent patterns that the events appear both in front and rear of comma. Besides extracting the pairs of the events, we identified the keywords that represented the relationships of the events, which positioned just in front of the comma. In Japanese language, these keywords have a role of conjunctions which connects (more than) two event descriptions.

PROPOSED METHOD EXTRACTION OF EVENTS AND CAUSALITIES

Events are phenomena in wear of machinery. We extracted the causality among the events from diagnostic reports by means of the morphological and dependency analysis[6]. Figure 1 shows the preprocessing steps. The events are obtained in their form of “subject + predicate”. In the first step, the nouns related to the events were extracted in advance, and were added to the noun list. The nouns indicated “the state”, “the material-substance”, “the parts” which appeared in the target diagnostic reports. In the nouns, the infrequent ones were variants

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Figure 1. Process of extracting the description features of causality(Japanese)

2.2

CLASSIFICATION OF EQUIPMENTS WITH SIMILAR CHARACTERISTICS

There are various types of lubricanting oil, equipments and wear particles generated by

17


Table 1. Example of extracted events with every semantic types .

Types Material Substance Parts State

Subject Iron Severe particle Gear Bearing metal Oil film shortage Temperature

Predicate is detected is detected is worn is worn occurs rises

them. The combinations of oil and equipments are also various. Even if experimental data of some fields are equal, wear conditions of machinery can be different depending on oil and equipments. In order to find relations between the wear conditions, oil and equipments, we, first, applied a clustering technique to find lubricating oil clusters, and classified equipments with similar numerical tends based on the oil clusters and other information, such as (rough) equipment categories, foreign substances from the outside, additives in lubricating oil and metals used as material of lubricated parts. We call the obtained clusters as equipment clusters. Since the number of equipment clusters were not clear, we used two-step clustering method to set the optimal number of clusters.

2.3

CLASSIFICATION SEVERITY

OF

WEAR

We generated a decision tree for each equipment cluster to extract rules in wear conditions of machinery. A target variable of the decision tree was wear severity of lubricated parts of machinery, which has the categorical values, “Severe”, “Partly severe” and “Not severe”. Input variables of the decision tree were the fields related to lubricating oil and equipments. Based on the generated decision tree, it was expected to find rules to predict wear severety with the values of input fields. In order to generate decision tree, we used C&R Tree, which is effective in classification by numerical value.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

2.4

GENERATION OF MODELS TO PREDICT WEAR CONDITIONS

In order to get rules that explain causality between causes and results of wear phenomena, we combined the conditions in the desicion tree to obtain wear severity and the causality extracted from diagnostic descriptions in Section 2.1. The prediction models were generated for each equipment clusters. The procedure to generate the model is as below: First, we extracted rules by following branches from the root to specific leaves of the decision tree. After that, we matched each of single predicates in the rules to the results in the causarity obtaied in Section 2.1. As a result, we expected to get prediction rules that relate the causes in the causarity (root causes) to wear severity of machinery (the target variable of the dicision tree). Second, we searched the result of the causarity whose cause matches a single predicate in the rules, which we call a predicted event. Combining the causarities with the mediation of the predicates provides us with another kind of rules, which relate the root causes to the predicted event. The causarities whose cause is the predicted event and whose result is the states of wear particles were joined to the rules to relate the root causes to the states, whose resultant rule we called condition rules. The condition rule makes it possible to diagnose, from the wear particle form, the background which led to the wear conditions. In order to diagnose a wear condition, the observed forms of wear particles can be input. If the input forms of wear particles correspond to the forms obtained from the condition rule, it is expected that the causarity in the condition rules occurs, which should be output as the causes of the current wear conditions. If the forms of wear particles are not input, all predicted events with their corresponding prediction rules are output as candidates of the events that will occur and their subsidiary information. Figure 2 illustrates prediction rules and condition rules obtained with the above procedure. Let us focus on the leaves of wear conditions reachable from a branch specified by “Viscos-

18


ity — low” and “Pollution level — high” and “Iron — high”, and let us assume that the causalities,“Oil+degradation” causes “low viscosity”, “Gear+wear” causes “high pollution level” and “high density iron particles”, “high density iron particles” causes “particle biting”, and “particle biting” generates “cutting particles” are obtained from diagnostic report. In this case, “Oil+degradation” and “Gear+wear” are found as the causes of the conditions appearing in the branch, including “high density iron particles”. Prediction rules are, therefore, that “Oil+degradation” and “Gear+wear” generate a severe wear condition. As predicted events, we find “cutting particles” that are caused by “high density iron particles” through “particle biting”. As a result, a condition rule is found as a flow from “Gear+wear” to “cutting particles”.

the causalities were extracted out of 22 causalities in the correspondence table, which suggests that 70% of extracted causalities are appropriate. The causes for normal wear particles could not be extracted as well as for most of pollution particles. The reason why we could not extract them will be that the description about well-known causalities or less frequent events might be omitted. Table 2. The extracted causalities of wear particles.

Particles form Normal Cutting Severe Fatigue Corrosion Erosion Metal bush Chemical change Pollution

3.2

Figure 2. Process of generating prediction models.

3 3.1

EVALUATION EXTRACTION OF EVENTS AND CAUSALITIES

We extracted the 1200 distinct causalities from 10,000 diagnostic reports with this method. In order to evaluate the extacted causalities, we prepared a table that summarized correspondence between wear particles and their causes. Then, we evaluated extracted causalities by comparing them with the correspondence table. Table 2 shows an example of extracted causalities. This table shows that 15 causes in

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Cause Not extracted Foreign substance bites Oil film severed Flaking occurred Acid was increased Not extracted Oil ring was worn Oil was polymerized Not extracted

CLASSIFICATION OF EQUIPMENTS BY CLASSTERING

Lubricating oils were classified into four oil clusters. Figure 3 shows their characteristics. The items in the figure are in the order of importance for classification. We used viscosity, oil code, and acid numbers, as the variables which the clustering algorithm was applied to. The horizontal axis of the graph is the value of the variables, and the vertical axis is the relative frequency of diagnostic cases. As can be seen from this figure, it was confirmed that both viscosity and acid number had low value in Cluster 1, both had from low to medium values in Cluster 2, viscosity had low value, and acid number had medium value in Cluster 3, both are high value in Cluster 4. The equipments were also classified into four equipment clusters based on oil clusters and fields, such as equipment categories, foreign substances from the outside, additives in lubricating oil and metals used as material of lubricated parts. Figure 4 shows characteristics of equipment clusters. Cluster 1 contained equip-

19


ments maintaining the fluid lubrication. Cluster 2 contained equipments under high pressure conditions using high viscosity of lubricating oil with additives, such as molybdenum. Cluster 3 contained low-degree polluted equipments using pressure oil, and Cluster 4 contained equipments frequently using additives to suppress acidic substances and the foreign substances. For classification, oil cluster, equipment class, equipment code, bearing division, BA, CA, MO, ZN were important items. Since equipment class, equipment code, bearing division are items that can classify equipments to some extent, they were the important items. In addition, from the order of items in Figure 4, BA, CA, MO, ZN used as additives in lubricating oil, and oil clusters were found to be the important items. These findings suggest that there is correspondence between classification of equipments and the additives in their oil.

Figure 4. Overview of equipment clusters.

Figure 3. Overview of oil clusters.

3.3

CLASSIFICATION OF WEAR SEVERITY BY DECISION TREE

We generated decision trees for each equipment cluster. These decision trees have its own number corresponding to the equipment cluster number. We got results that wear severity was roughly classified by density of iron particles and pollution levels of lubricating oil in all of decision trees. However, branch conditions were different in the respective dicision trees if the branch is near a leaf node. Wear severity was classified by the items related to the properities of lubricating oil such as viscosity, acid number, pollution level and barium in Decision tree 1, by iron and viscosity in Decision tree 2, by items related to pollution of

ISBN: 978-1-941968-26-0 ©2015 SDIWC

the lubricating oil such as counting pollution level and silica in Decision tree 3, and by items lead to degradation of the lubricating oil such as acid number and water in Decision tree 4. These findings suggest that the influence of the particles in lubricating oil was larger than the properties of oil and additives which was important at classification of equipments for wear severity. Table 3 shows the results of the accuracy analysis for each decision tree. The accuracy analysis calculated the accuracy rate of predicting target variable by analysis data that was not used at generation of decision tree. Its results suggest that Decision tree 3 had the lowest predictive accuracy, whose reason is that the clustering of equipments could be inadequate because of assignment of equipment class was insufficient.

20


Table 3. The results of accuracy analysis for each decision tree.

cases in Model 4 were less than other models.

Decision tree Prediction accuracy 1 71.47% 2 70.35% 3 66.31% 4 73.94% 3.4

MODELS TO CONDITIONS

PREDICT

WEAR

We generated four prediction models by combining extracted causalities with each decision tree. The prediction model number correspond to the decision tree number. Figure 5 shows Model 2. In order to evaluate the models, we compared wear conditions of analysis data predicted by models with the contents of diagnostic report for its analysis data. Their diagnostic reports and analysis data were used to prepare 10 evaluation data for each model. In the evaluation, we evaluated whether contents of diagnostic report and diagnosis by model are perfectly matched, partially matched and mismatched. Table 4 shows the results of evaluating Model 2, where we used the score 1, 0, -1 for perfect match, partial match and mismatch respectively. As can be seen from this table, if the prediction by decision tree was wrong such as Data 3,4 and 8, proper wear condition was not obtained. The reason why calcium was mixedin was specifically described in diagnostic report of Data 4. However, since Decision tree 2 has no branch about calcium, diagnosis for this wear condition was not obtained. Since Leaf 1 was only condition appearing in the branch, “high density iron particles”, data classified into Leaf 1 are mostly classified into partial matched. Table 5 shows results of evaluation to all models. In the case of mismatch, if the prediction by decision tree was wrong, the models could not get proper wear condition. In the case of partial match, the models could not cover all of wear conditions described in the diagnostic reports. Because of these reasons, perfect match

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Figure 5. Generated Prediction model 2.

Table 4. The result of Model 2 evaluation.

Data Leaf 1 1 2 1 3 4 4 3 5 4 6 1 7 4 8 5 9 2 10 1

Wear severity Evaluation Yes 0 Yes 0 No -1 No -1 Yes 1 Yes 1 Yes 0 No -1 Yes 1 Yes 0

Table 5. The result of evaluation for all obtained models, which shows the number of cases perfect match/partial match/mismatch for the models.

Model 1 2 3 4

4

Perfect match 3 3 3 1

Partial match Mismatch 7 0 4 3 3 4 4 5

CONCLUSIONS

In this study, we generated the models to diagnose equipment wear conditions using lubricating oil analysis. The models were generated by combining the causalities extracted from description in diagnosis reports and the rules involved in wear severity of machinery. First, in order to extract the causalities from diagnostic reports, we extracted the events and

21


the description features of causality between the events in advance. We extracted events, the combinations of words, such as “the materialsubstance + its change/detection”, “the parts + its state”, “the state + its extent/change” , which are the combinations of “ the modified source + the modifying destination” out of the sentences in the descriptions. Since the causalities between extracted events frequently straddle breaks in sentences, we extracted causalities that appeared around commas breaking sentence. In order to evaluate the result, we closely confirmed the obtained causalities and found that 15 causalities (70%) out of 22 obtained causalities were correct. Second, in order to extract the rules for wear severity of machinery, we applied data mining to analysis data. We employed a clustering technique to find clusters of equipments, and clarified that properties of oil and additives were the important items at classification, which suggested that there was correspondence between classification of equipments and the additives in their oil. We generated decision trees, where target variable is wear severity for each cluster. As the results of accuracy analysis for each decision tree, the lowest predictive accuracy was 66.31%, and the highest predictive accuracy was 73.94%. The influence of the particles in lubricating oil was larger than the properties of oil and additives which was important at classification of equipments for wear severity. Finally, we generated prediction models by combining extracted causalities with each decision tree. In order to evaluate the models, we compared diagnostic reports with wear conditions obtained by models using 10 evaluation data. Although many of each content was partially matched, there was no mismatch in the model for equipments maintaining fluid lubrication. Since the models realized simple diagnosis, we showed the possibility of automatic equipment diagnosis for lubricating oil analysis. In the future tasks, it will be necessary to take account of the probability or weight of the events in order to extract the detailed wear con-

ISBN: 978-1-941968-26-0 ©2015 SDIWC

ditions of machinery. Moreover, since we extracted only one-to-one causalities between the events in this study, it will be necessary to extract the co-occurrence relation among events or the result occurred by multiple causes. ACKNOWLEDGEMENTS I wish to thank Triobtex Inc. for providing lubricating oil analyzed data that we used in this study. REFERENCES [1] M. Muraki.(2007). Schematic science of tribology friction and lubrication technology. Nikkan Kogyo Shimbun, Ltd, Japan, 1st edition. [2] TRIBOTEX INC, TRIBODIAGNOSIS, http://www.tribo.co.jp/tribodiagnosis.html, 2015 [3] N. Omori and T. Mori. (2012). Proposal of method extracting causality about products and parts from the failure case document. pages 1192-1195. Language Processing Society announced papers. [4] M. Ishida.(2008). Companion to Text Mining by R. Morikita Publishing Co., Ltd., Japan, 1st edition. [5] M. Berry and G. Linoff. (1999). Data Mining Techniques: For Marketing, Sales, and Customoer Relationship Management. Kaibun-do, Japan, 1st edition. [6] T. Kudo, Y. Matsumoto. (2002). Japanese dependency analysis using cascaded chunking. In CoNLL 2002: Proceedings of the 6th Conference on Natural Language Learning 2002 (COLING 2002 PostConference Workshops), pages 63-69.

22


‘Fuzzy’ vs ‘Non-Fuzzy’ Classification in Big Data

1,2

Malak EL-Bakry1, Soha Safwat 2 and Osman Hegazy3 Faculty of Computer Science, October University for Modern Sciences and Arts, Giza, Egypt 3 Faculty of Computers and Information, Cairo University, Giza, Egypt 1 [email protected], [email protected],[email protected]

ABSTRACT Due to the huge increase in the size of the data it becomes troublesome to perform efficient analysis using the current traditional techniques. Big data puts forward a lot of challenges due to its several characteristics like volume, velocity, variety, variability, value and complexity. Today, there is not only a necessity for efficient data mining techniques to process large volume of data but also a need for a means to meet the computational requirements to process such huge volume of data. The objective of this research is to compare fuzzy and non-fuzzy algorithms in classification of big data, and to provide a comparative study between the results of this study and the methods reviewed in the literature. In this paper, we implemented the Fuzzy K-Nearest Neighbor method as a fuzzy technique and the Support Vector Machine as nonfuzzy technique using the map reduce paradigm to process on big data. Results on different data sets show that the proposed Fuzzy K Nearest Neighbor method outperforms a better performance than the Support Vector Machine and the method reviewed in the literature.

KEYWORDS Big data; Classification; Fuzzy K Nearest Neighbor; Support Vector Machine; Hadoop; MapReduce

1 INTRODUCTION Various innovations of technology are driving the spectacular growth in data and data gathering, this is the reason behind the question of “why big data has become a recent area of strategic investment for its organizations?” [1]. Big data is a group of enormous volume of structured and unstructured data from different

ISBN: 978-1-941968-26-0 ©2015 SDIWC

sources. There are different sources of big data such as data coming from social network, data generated by machine, and traditional enterprise. Big Data is huge and difficult to develop using ordinary database and software techniques, and due to its complexity it require a new architecture, techniques, algorithms, and analytics to manage it and read out the values and extract the hidden knowledge from it [1]. Now it is impossible for analysts to extract a meaningful useful conclusion from data in a short time frame due to the huge volume of the data. So techniques of data mining are looked upon as tools that can be used to automate the process of knowledge discovery and define relationships and patterns of likeness given a completely random and raw data set. Unsupervised data is the majority of the data collected for analysis. This shows that there is a need for an effective technique that can process on such unsupervised data sets and switch what might seem to be totally random and meaningless into something more meaningful and valuable [2]. Accessing data, computing data, domain knowledge, privacy of data, and data mining are main problems of the big data [3]. Due to these challenges, processing data and data mining techniques became a critically important role in development technology [2]. Classification is one of the most useful techniques in data mining that classifies data into the structured class or groups and helps the user in discovering the knowledge and future plan. Classification supplies the user with an intelligent decision making. Classification consists of two phases; the first phase is learning process phase in which a huge training

23


data sets are analyzed, and then it creates the patterns and the rules. The second phase is evaluation or testing, it records the accuracy of the performance of classification patterns. The purpose of classification is to be able to use its model to predict the class label of objects whose class label is unknown. Various forms can be represented; For instance, neural networks, classification rules, mathematical formulae or decision tree [1]. Support Vector Machine (SVM) is the classification technique which can be applied to process on large data. The complex and big data can be left to the SVM because of its results when it is greatly affected with too much noise in the datasets. Over fitting problem can be solved with an optimized SVM algorithm because of its effectiveness in classification. SVM can make use of certain kernels to uncover efficiently in quantum form the largest Eigen values and corresponding Eigen vectors of the training data overlap (kernel) and covariance matrices [3]. K Nearest Neighbor (KNN) is one of the most standout classification algorithm in data mining, which is based on homogeneity, which implies drawing a comparison between the given test record with training records which are similar to it. K Nearest Neighbor classification provides us with the decision outline locally. It was developed due to the need to carry out discriminate analysis when reliable parametric estimates of probability densities are unknown or hard to define. K is a constant pre-defined by the user. The testing data are classified by giving the label which is most frequent repeated among the k training samples nearest to that query point [4]. Fuzzy logic is a technique of computing which is based on the "degrees of truth" not like the traditional "true or false" (1 or 0) techniques which are the basics of the modern computer. The first one who invented the idea of fuzzy logic was Dr. Lotfi Zadeh from the University of California at Berkeley in the 1960s. Solving the problem of the computer understanding of natural language was the problem that leads Dr.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Zadeh to think in the fuzzy logic. Absolute terms (0 and 1) are very difficult to use in describing the natural language. It is very difficult to translate the natural language to 0 and 1. Fuzzy logic includes two extreme cases of the truth which are zero and one. In addition, it includes the different cases of truth in between. For example, the output of a comparison between two things could not be "tall" or "short" but ".38 of tallness" [5] [14]. The fuzzy k-NN classifier works by assigning a membership value to the unlabeled signature that supply the framework with proper data for estimating the certainty of the decision. Each of the defined classes has a fraction of unlabeled signature defined by fuzzy membership coefficient. Delineation means that the membership of two classes is relatively high, while confusion means that the membership between two classes in very low. When assigning a non-fuzzy label to the signature the above data becomes very important [6]. Assigning an object to the unknown class is a greater advantage of using a fuzzy system over crisp system. On the other hand, crisp system can assign the object to wrong class. The fuzzy k-NN classifier classifies the data according to the training data and the following fuzzy information taken from it. The user can indirectly control the defuzzification level to specify the percentage of wrong decision is “worth” to the process. In a lot of cases, setting the defuzzification level has more advantage than not setting it; because when more defects are categorized as unknown is much better than classifying it wrongly. This is so accurate in many cases where classifying a defect wrongly could result in a stronger economic effect or income loss [6]. Several classification techniques using Map Reduce architecture are implemented as the use of linguistic Fuzzy rule by Victoria Lopez, Sara Del Rio, Jose Manuel Benitez and Francisco Herrera [7], the k-Nearest Neighbor Algorithm by Prajesh P Anchalia and Kaushik Roy [2], the Support Vector Machine by Ke

24


Xu, Cui Wen, Qiong Yuan, Xiangzhu He, and Jun Tie [8], and Neural Networks by Chetan Sharma [9]. In this paper we implemented the Fuzzy K Nearest Neighbor method and the Support Vector Machine using the map reduce paradigm to process on big data to enhance the performance of several methods that have been reviewed in the literature [7]. To do so, we rely on the success of the Map Reduce framework. The map phase performs the splitting of the data, and the reduce phase perform join to all outcome from the mappers and gives us the final output. To test the performance we conducted experiment on different data sets. The experimental study indicates an analysis of the accuracy of the testing. The rest of this paper is organized as follows: section II briefly introduced the big data classification; section III contains the approaches developed in this work. The proposed system is shown in section IV. The experimental results and analysis are then discussed in section V. Finally section VI summarizes and concluded the work. 2 BIG DATA CLASSIFICATION 2.1 Big Data Big Data definition depends on who is describing it, and in which context. So there is no comprehensive definition for it. However, at common level, almost every definition of big data summarizes the concept to huge and increasing data masses and the process of analyzing that data. The basis of competition, productivity enhancement and creating important value for the world economy by decreasing waste and expanding the quality of items and services will become the main reason behind the use of Big Data in analyzing and making better decisions. The big data has three characteristics which are the three Vs. The three Vs are: Velocity, Variety, and Volume. Volume defines the large quantity of data produced periodically. Velocity is the speed at which data

ISBN: 978-1-941968-26-0 ©2015 SDIWC

is generated. It should have high rapidity data. Variety is the various data formats, types, and structures. The type of data may cover various varieties such as Text, numbers, photos, video, audio, time series, and social media data. It may also include static data and dynamic data. Collecting many types of data can be generated by a single application. All these types of data need to be attached together to extract the knowledge. Extracting information will need advanced method to process due to its high volume, velocity, and variety [1]. Medicine, Physics, Simulation, RFID, Astrometry, Biology and a lot more are different applications for big data. There are various types of big data such as structural, relational, textual, graph data, streaming, semi structured [1]. The problem of meeting computational demand is posed by the big data. This problem is discussed professionally by distributed computing. Both hardware and software components deal with distributed computing. These components are physically separated but communicate with each other and work like a single system on the whole. So, we split the processing of massive volume data among many computers that are networked and work together towards a common goal. Distributed computing environment basically consists of computers connected over a network and interact with each other constantly to achieve a common goal. Individually, each computer connected on the network is called a node, and collectively all nodes form a cluster [2] [15]. The technology needs new architecture, algorithms, techniques, and technical skills for handling the big data. Therefore to deal with big data, there is a need for experts. Enterprises face a lot of challenges in handling Big Data and capturing information from it. The format of the data, type of analysis, and processing techniques play a role in classifying the big data [1].

25


2.2 Data Mining Data mining is a method that generates and detects the relevant patterns and relationships of a dataset. Its main objective is to define approaches and algorithms to browse the given data set and predict the data direction. Data Mining is not essentially a new method. Similar manual approaches have been used by statisticians to review data and detect its direction. The only thing that has been changed is the volume of data today. The volume of the data today has been increased extremely, and this gives a rise to automated data mining techniques that investigate data direction rapidly. Determining the outcome of the data analysis by the users can be done by the parameters they choose, so flexibility became one of the advantages of automated data mining [2]. Data mining can be achieved in two ways either supervised or an unsupervised way. Supervised learning is learning through a system which receives a dataset and decisions. Then it includes some mathematical functions that map the input to the output and find what decision it supposed to be. On the other hand, the unsupervised learning means that we have to group the data without referring to any type of predefined cases [1].There are various methods of data mining system which can be implemented with big data. The prime methods used with data mining are: classification; Evolution analysis; Outlier Analysis; and Cluster analysis. Each one of them will be discussed below. First, the classification; which is the technique of detecting a function or model that characterizes data classes for the purpose of being qualified to use the model to guess the class of record whose class label is unidentified. Analyzing training data is the derived model which can be symbolized in many forms; such as decision tree, classification rules, mathematical methods or neural networks. Second, the evolution analysis; which is based on time series data of previous years. Predicting future direction in stock market is done

ISBN: 978-1-941968-26-0 ©2015 SDIWC

according to the regularities in such time series, and contributing to decision making. Third, the outlier analysis; which is based on statistics and probability that estimates a probability model for the data or uses distance measures in which far objects from any other cluster are considered outliers. Fourth, the cluster analysis; where the class labels are not available in the training data sets. Class labels are generated using these techniques. Data sets in the cluster are placed in groups depending on their likeness. The most important clustering methods include hierarchical methods, portioning methods, model based methods, density based methods, and constraint based clustering method [1]. We choose to implement the classification due to its efficiency in data analysis which can be used in extracting models that describe the important data classes or guessing the upcoming data direction. This analysis gives us a better understanding of the data at large [10]. Classification algorithms typically contain two phases: • Training Phase: in this phase, a model is structured from the training sample. • Testing Phase: in this phase, the model is used to specify a label to an unlabeled test sample. 2.3 MapReduce Paradigm MapReduce is published by Google as a programming model for conducting difficult combination of huge amount of data. [1] The aim of the MapReduce program is to process on massive volume of data on multiple machines. MapReduce divides data into independent blocks. Due to this division, a single task is divided into multiple subparts. Each one is handled by an independent node [11]. MapReduce are divided into two different steps; Map and Reduce. Each step is done parallel on sets of pairs [2]. The Map function takes the divided data values as input, then it performs any function to every value in the input set and produces an

26


output set. The output of the map is in the form of pairs stored on each node. The output of the Map function is stored and sent to the Reducer [2]. The Reduce function takes the output from the Map function as an input for it, and then generates another set of as final output. The Reducer cannot start until all the Map stage is finished and the results are sent to the appropriate machine. The MapReduce framework consists of a single Master Job Tracker and multiple Task Trackers. The task tracker can be any node in the cluster. The Master Job Tracker is responsible for division of the input data, task scheduling, failure of the machine, re-execution of un-succeeded tasks, inter-machine communications and task status monitoring. The task tracker is responsible for executing the tasks assigned by the master. There is a file system to store both input and output files. The single Job Tracker can be a single point failure in this framework. MapReduce is an appropriate technique to deal with huge datasets, and therefore ideal for mining Big Data of petabytes size that do not fit into a physical memory [2]. The MapReduce algorithm is implemented using many commercial and open-source technologies as a part of their internal architecture. One of the most efficient and popular techniques of implementing the MapReduce is the Apache Hadoop, which aims to be used for data executing in a distributed computing environment. Any programming language can be used in implementing the MapReduce algorithm [11]. 2.4 Hadoop ApacheTM Hadoop is an open source application that runs on a distributed computing environment and supports processing of Big Data with huge volume. It has been evolved from The Google File System [2].The architecture of the Hadoop contains many components; such as, the master server which directs jobs to the machines of the underlying

ISBN: 978-1-941968-26-0 ©2015 SDIWC

worker. It also contains a package of components usually named the “Hadoop Common Package.” This package consists of components. For example, the “Hadoop Distributed File System” (HDFS); MapReduce engine; and scripts to run the Hadoop installation [11]. HDFS contains name nodes and data nodes; they are downloaded on the server of the master and workers separately. The name node is responsible for managing data using data nodes located in the machines of the workers and mapping those files to data nodes. The data nodes on the worker computers implement read and write requests as required [2]. Data in the Hadoop framework is saved across multiple nodes in the HDFS. Replication of data in HDFS is done three times on separate nodes to provide protection from failure. To insure data integrity, a checksum for the data blocks is continuously calculated. The programming libraries are used to perform the distributed computing tasks, which implement the MapReduce algorithm. All of these components work with each other to process on Big Data in a batch processing mechanism [11]. Due to the characteristics of Hadoop, it achieves its goal of processing by breaking down the given data set and process on it individually at separate nodes that are connected together. Data will be subject to failure because of the distribution of it on multiple nodes. So, when a failure is detected the process can be restarted automatically again on another node. It also creates a copy of missing data from the available replicas of data. There is a failure point when only one single Namenode is available [2]. There are numerous open source projects based on the top of Hadoop; such as, Hive; Pig; Mahout; and Pydoop. Hive is a data warehouse framework that analyzes complex data. Pig is a dataflow framework that produces a series of MapReduce programs. Mahout is responsible for the machine learning libraries that focus on clustering, classification, frequent item-sets mining and evolutionary programming. Pydoop

27


is a python package that supports API for Hadoop MapReduce and HDFS [2]. 3 PRELIMINARIES 3.1 Support Vector Machine Support Vector Machines is a technique of the supervised learning methods used for classification. The aim of SVM is to find the best partition hyper plane by maximizing the margin between the two classes. The data instances are arranged in a way where there is a hyper plane that divides it from the others by applying the kernel equations. Kernel equations may be linear, Gaussian, quadratic, or any other equation that reach the same purpose. The purpose of the kernel equation is to convert the linearly non-separable data in one domain into another domain where the instances become linearly separable [3]. When we succeeded in splitting the data into two different classes, then our goal is to get the best hyper-plane to divide the two types of instances. Deciding the target variable value of the upcoming prediction is the job of this hyper-plane. We should choose a hyper-plane that increases the margin between the support vectors on the both side of the plane. One of the most important things about Support Vector Machines is that the data to be divided must be binary. And in case if the data is not binary, Support Vector Machines deals with it as it is, and performs the analysis through a series of binary assessments on the data [4]. Algorithm 1" Support Vector Machine (SVM) algorithm" 1: Finding Pair of Points that are closed Candidate SV ={closest pair from classes that are opposite} do Find a violator 2: Adding a sample to the Support Vector data set candidateSV = candidateSV violator 3: Pruning

ISBN: 978-1-941968-26-0 ©2015 SDIWC

if any < 0 as a result of adding c to S then candidateSV = candidateSV \ p repeat till all such points are pruned end if while there are violating points

3.2 Fuzzy K-Nearest-Neighbor A fuzzy KNN classifier was designed by Keller et al. [12], where class memberships are given to the sample, as a function of the sample‟s distance from its k nearest neighboring training samples. A Fuzzy K-NN Classifier is one of the most successful techniques for applications due to its simplicity and also because of giving some information about the certainty of the classification decision. Keller et al assume that the improvement on the error rate might not be the major advantage from using the FKNN model. More importantly, the model offers a percentage of certainty which can be used with a”refuse-to-decide” option. Thus objects with overlapping classes can be detected and processed individually [12]. Algorithm 2 “ Fuzzy K Nearest Neighbor Classifier 1:For i=1 to m do Compute distance d( X: Training data X: unknown sample 2: Select K smallest distances 3: Assign an input a membership vector “Soft labels”.

[

{

]

( ) ( )

( ) ( )

}

4: Calculate Membership function

∑ ∑ 5: Return

max

28


4 PROPOSED SYSTEM

4.1 Big Data Classification Using FKNN

In this section we show the details for the proposed method to classify big data using Fuzzy K Nearest Neighbor and Support Vector Machine algorithms.

After finishing organizing our training and testing data, we can apply the Map Reduce Fuzzy K Nearest Neighbor technique in a distributed environment by the following algorithm discussed below.

We divide the data sets to training data sets and testing data sets. The training data sets are 75% of the whole data, and the rest 25% of the whole data are the testing data sets. MapReduce divides data into independent chunks and the size of division is a function of the size of data and number of nodes available. As shown in Fig. 1 the data sets are divided on several mappers using the map function; each mapper contains the same number of samples. The reduce part takes the results of individual mappers and combines them to get the final result. The idea of the model is to build individual classifier on each group and use each classifier to classify the testing data and send the class label to the reducer function, then the reducer take the majority vote to decide the final class label for the testing data. Mapper

𝑴𝒌

FKNN/SVM

CLASS LABEL

FKNN /SVM

CLASS LABEL

FKNN /SVM

CLASS LABEL

REDUCER MODE (Class label)

FINAL OUTPUT

Figure 1: The Proposed System

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Algorithm 3 :Map Reduce design for FKNN Read the Training data Load TrainingData Procedure FKNN MapDesign 1:Load testing data file Load TestFile 2: Create a matrix contains testing data TestData = TestFile 3:Read Samples from TestData one at a time While Not End Of File 4: Read Mappers from TrainData one at a time While Not End Of Mappers 5: Call FKNN algorithm Result = FKNN(TrainingData, TestData ,K ,M) 6: Send Result to the Reducer Design End While End While call reducer End procedure Procedure FKNN ReducerDesign Load Value of K Load value of M 1:Load testing data file TestData = TestFile 2: Create a Vector contains testing data class label 3: Create a Matrix Outputs , M is number of mappers and D is number of samples in TestFile 4: Initialize Outputs Matrix for all class labels SET counters to ZERO Outputs =zeros(D,M) 5: Read Result from MAP functions one at a time while NOT END OF mappers 6: Write The output from MAP function into Matrix output(i)= Result 7:Assign the class label with the highest count for the testData sample SampleOutput= mode(output)

29


Mode: is the majority vote End While End procedure

As shown in the Map Reduce design for FKNN, Map and Reduce are two different steps. Each step is done parallel on sets of pairs. So the programs are bifurcated into Map and Reduce stage. Each Mapper function takes its training data after dividing it on the set of mappers, and also takes the testing data. The Map routine performs the FKNN function of training which is calculating the distance of each data point with the classes, then list out the class of the unknown data. After that the class label is sent to the reducer function by all mapper, and then the reducer function uses the majority vote function to classify the testing samples. 4.2 Big Data Classification Using SVM After organizing our training and testing data, we can apply the Map Reduce Support Vector Machine technique in a distributed environment by the following algorithms discussed below. Algorithm 4:Map Reduce design for SVM Read the Training data Load TrainingData Procedure SVM MapDesign 1:Load testing data file Load TestFile 2: Create a matrix contains testing data TestData = TestFile 3:Read Samples from TestData one at a time While Not End Of File 4: Read Mappers from TrainData one at a time While Not End Of Mappers 5: Call SVM training algorithm ClassLabel=SVM(TrainingData, TestData) 6: Send Result to the Reducer Design End While End While call reducer End procedure Procedure SVM ReducerDesign

ISBN: 978-1-941968-26-0 ©2015 SDIWC

1:Load testing data file TestData = TestFile 2: Create a Vector contains testing data class label 3: Create a Matrix Outputs , M is number of mappers and D is number of samples in TestFile 4: Initialize Outputs Matrix for all class labels SET counters to ZERO Outputs =zeros(D,M) 5: Read Result from MAP functions one at a time while NOT END OF mappers 6: Write The output from MAP function into Matrix output(i)= Result Assign the class label with the highest count for the testData sample SampleOutput= mode(output) Mode: is the majority vote End While End procedure

As shown in the Map Reduce design for SVM, Map and Reduce are two different steps. Each step is done parallel on sets of pairs. Each Mapper function takes its training data after dividing it on the set of mappers, and also takes the testing data. The Map routine performs the Support Vector Machine algorithm for training; in the training we use the RBF kernel function. After finishing the training phase we read testing data, it starts by reading sample at a time and classify this sample individually to each classifier, then send the class label for each mapper to the reducer function. Then the reducer function uses the majority vote function to classify the testing samples. 5 EXPERIMENTAL RESULTS In this section, we describe the results obtained from classification using Support Vector Machine (SVM) and Fuzzy K Nearest Neighbor (FKNN). 5.1 Data Sets In this experimental study we will use big classification data sets taken from the UCI

30


repository. Table1 summarizes the main characteristics of these datasets. For each dataset, we show the number of records, number of attributes and number of classes [13]. Table 1: Data Sets Description Data Covtype Covtype-2 Poker Poker-2

No of records 581012 581012 1025009 1025009

No of attributes 54 54 10 10

No of classes 7 2 9 2

5.2 Evaluation Method The purpose of classification is to build a classifier from given data and predict the future data correctly. The most commonly used performance measurement is the classification accuracy. For a given finite number of records, the empirical accuracy is defined as the ratio of the correct classification to the number of given records. Empirical accuracy=

(1)

5.3 Experimental Results Firstly, we compare the performance of the Support Vector Machine and the Fuzzy K Nearest Neighbor. We used the SVM and FKNN algorithm for four data sets. In the SVM the parameter used is the kernel function, and in FKNN the parameters used is the K and the µ, the K is the user defined constant and the µ is the fuzzification value. In this experiment in the SVM, we used the RBF as the kernel function for all datasets. In the FKNN, we tried a lot of K and the best outcome comes when we used the K with value 50 and µ with value 0.6 with the poker data sets and K with value 5 and µ with value 0.6 with the Covtype datasets. After finishing classification we calculate the accuracy for the testing phase.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Table 2: Accuracy Results of Proposed System Classifier Data Covtype Covtype-2 Poker Poker-2

SVM

FUZZY KNN

41.221 66.4234 45.231 53.02

54.7694 75 62.1694 69.0631

According to table 2, we can see that the FKNN classifier gave higher accuracy comparing to the results when we used the SVM. Second, we compare the performance of the proposed system with a paper reviewed in the literature review which used linguistic fuzzy rule [7]. Table 3: Comparison of Accuracy Methods

SVM

FKNN

FRBCS

Data Sets Covtype-2 Poker-2

66.4234 53.02

75 69.0631

74.96 60.35

100 90 80 70 60 50 40 30 20 10 0

SVM FKNN FRBCS

Covtype-2

Poker-2

Figure 2: Comparison between the proposed system (SVM and FKNN) and literature reviewed FRBCS

As shown in table 3 and figure 2, we compare the results of the proposed model and a method mentioned in literature on the same data set, the proposed model gave better performance while using FKNN than that of the SVM and the FRBCS.

31


6 CONCLUSIONS In this study, we introduced a comparative study of classification on big data; we used two classification algorithms; the Fuzzy K Nearest Neighbor and the Support Vector Machine using the MapReduce paradigm. The two proposed algorithms consist of two parts; the mapper and the reducer. The mapper algorithm is used to divide the data sets in to chunks over the computing nodes and produce a set of intermediate records. The records produced by the map function take the form of a “(key, data)” pair. Mapper in the individual nodes execute the computing process and send the results to the reduce function. The reducer algorithm receives the results of individual computations and put them together to obtain the final result. Good accuracy of the performance was obtained using the Fuzzy K Nearest Neighbor method than the Support Vector Machine and the Fuzzy Rule based classification system reviewed in the literature.

[6]

[7]

[8]

[9] [10]

[11] [12]

[13]

[14]

REFERENCES [15] [1]

[2]

[3]

[4]

[5]

T. Smitha, Mca, M. Phil, M. Tech, V. Kumar „Application of Big Data in Data Mining‟. International Journal of Emerging Technology and Advanced Engineering, vol. 3, no. 7, pp. 390-393, 2013. P. Anchalia, and K. Roy. „The K-Nearest Neighbor Algorithm Using Mapreduce Paradigm‟. Fifth International Conference, 2014. P. Koturwar, S. Girase, D. Mukhopadhyay, „A Survey of Classification Techniques in the Area of Big Data‟, 2015. S. Pakize, A. Gandomi, „Comparative Study of Classification Algorithms Based On MapReduce Model‟. International Journal of Innovative Research in Advanced Engineering (IJIRAE), vol. 1, no. 7, pp. 251-254,2014. X. Wu, X. Zhu, G. Wu, W. Ding, „Data mining with big data‟, IEEE Trans. Knowledge Data Eng. IEEE Transactions on Knowledge and Data Engineering, vol. 26, no.1, pp. 97-107, 2013.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

K. Tobin, S. Gleason, T. Karnowski, „Adaptation Of The Fuzzy K Nearest Neighbor Classifier For Manufacturing Automation‟. S, Río, V. López, J. Benítez, F. Herrera, „A MapReduce Approach to Address Big Data Classification Problems Based on the Fusion of Linguistic Fuzzy Rules‟, International Journal of Computational Intelligence Systems, 2015. K. Xu, C. Wen, Q. Yuan, X. He, J. Tie, „A MapReduce based Parallel SVM for Email Classification‟, Journal of Networks JNW, 2014. C. Sharma, Big Data Analytics Using Neural networks, 2014. N. Jain, „Data Mining Techniques: a Survey Paper‟, International Journal of Research in Engineering and Technology IJRET,vol. 2, no. 11, pp.116-119, 2013. B. Bhagattjee, „Emergence and Taxonomy of Big Data as a Service‟, 2014. J.M. Keller, M.R. Gray, J.A. Given, „A Fuzzy K Nearest Neighbor Algorithm‟, vol.15, no. 4, pp. 580585, 1985. I. Triguero, D. Peralta, J. Bacardit, S. García, F. Herrera, „MRPR: A MapReduce solution for prototype reduction in big data classification Neurocomputing‟, vol. 150, pp. 331-345, 2015. A. Sharma, B. Padamwar, „Fuzzy Logic Based Systems In Management And Business Applicaions‟, International Journal of Innovative Research in Engineering & Science, vol. 1, no. 2, pp. 1-6, 2013. Juniper Networks, Introduction to BIg data: Infrastructure and Networking Considerations, 2012.

32


On Definition of Automatic Text Summarization Pashutan Modaresi Stefan Conrad Heinrich-Heine-University of Düsseldorf Institute of Computer Science, Düsseldorf, Germany {modaresi, conrad}@cs.uni-duesseldorf.de ABSTRACT Research in the continuously growing field of automatic text summarization is branched into extractive and abstractive approaches. Over the past few decades, major advances have occurred in extractive summarization and a smooth transition from extractive to abstractive approaches can be observed in recent years. Despite advances, a proper definition of automatic text summarization has been mainly neglected by researchers. In this work we emphasize on the importance of an appropriate definition of automatic text summarization. We review previous definitions on text summarization, investigate their properties and propose our own definition.

KEYWORDS Text Summarization, Scientific Definition, Content Selection, Readability, Text Mining

1

INTRODUCTION

Modern research on automatic text summarization began almost 60 years ago with the work of Luhn [1] on automatic creation of literature abstracts. Over the years, much progress has been done in the development of algorithms to automatically summarize documents. Among the two major approaches of extractive and abstractive summarization, the first one has been investigated extensively in the literature. Despite satisfactory results in extractive summarization, researchers are focusing more and more on abstractive summarization in the recent years. Extractive summaries are usually created by concatenation of fragments of the source document with the addition of some post-processing. On the other hand, abstractive summaries are the result of rewriting or paraphrasing the source document, where a oneto-one mapping between the sentences of the

ISBN: 978-1-941968-26-0 ©2015 SDIWC

source and target document is not always possible. Abstractive summarization is considered to be the natural way of summarizing performed by humans, which is one of the reasons for its popularity in the recent years. As in other scientific disciplines, the first step to approach the automatic summarization problem is to define the problem itself. Previous research on extractive text summarization has revealed a disagreement in the community regarding the understanding and definition of the summarization problem. This can be acknowledged by the discrepancy between the various definitions of the problem proposed in the literature so far. In like manner, a similar flaw can also be observed in the field of abstractive text summarization. With this in mind, a need for a proper definition of automatic text summarization is being felt in the research community. This has been mainly neglected by researchers and consequently led to inconsistent foundations of this field. By inconsistent we mean that there is no single definition of a summary which has been agreed upon by researchers. We impose several requirements on a proper definition of automatic text summarization: • Universality: The definition should be valid for the known types of automatic text summarization. This includes indicative [2] and informative (according to functionality), single- [3] and multi-document [4] (according to input cardinality), hierarchical and flat [5] (according to output cardinality), extractive [6] and abstractive [7] (according to type), as well as generic, update [8] and query-guided (according to context). • Generality: The definition should not apply any restrictions on the implementation details of the various stages of automatic

33


text summarization. This includes particularly the representation type of the source document, content selection, scoring, lexical selection or text realization. • Minimality: The definition should be minimal, meaning that only a minimal number of properties or characteristics to reconstruct a text summary should be mentioned. • Exclusivity: The definition should be exclusive, meaning that the definition allows degenerate cases that one may wish to exclude. • Repeatability: The definition should be repeatable, meaning that applying the definition to a summary document as the input document, should either return a valid summary document, or prevent us from creating a new summary document if it is not possible. In Section 2 we provide an overview of the existing definitions of automatic text summarization and investigate their properties. A commonly used concept in automatic text summarization is compression rate. The usage of this concept will be criticized in Section 3 and the concept of readability will be suggested as a substitution. Content selection as an important part of any summarization system will be discussed in Section 4. In Section 5 we propose our own definition of automatic text summarization and finally in Section 6 we will conclude our work. 2

RELATED WORK

Various definitions of automatic text summarization have been proposed in the literature. Despite some commonalities, these also include contradictions in some cases. Furthermore the proposed definitions are mostly applicable to a certain type of automatic text summarization and lack the properties introduced in Section 1. The lack of a proper definition of automatic text summarization can be due to a conservative attitude in the community, as Das and Martin state: “. . . it seems from the literature that

ISBN: 978-1-941968-26-0 ©2015 SDIWC

any attempt to provide a more elaborate definition for the task [ of automatic text summarization ] would result in disagreement within the community”[3]. In this section we investigate the previous definitions of automatic text summarization proposed in the literature. We study their flaws and inspect the properties that make them inappropriate definitions of automatic text summarization. In fact Luhn did not propose a definition of text summarization, but rather he mentioned the purpose of a summary in the context of literature abstracts as: “the purpose of abstracts in technical literature is to facilitate quick and accurate identification of the topic of published papers. The objective is to save a prospective reader time and effort in finding useful information in a given article or report”[1]. Although this cannot be considered as a definition of automatic text summarization, but Luhn’s statement points to two important properties of a text summary. The first property is that the time and effort for reading a summary should be less than the one being consumed in reading the original document, and the second property is that a summary should accurately reflect the topic of the original document. In 1995 Maybury defined an effective summary as “[ a text that ] distills the most important information from a source (or sources) to produce an abridged version of the original information for a particular user(s) and task(s).”[9]. By mentioning that a summary is produced from a source (or sources), Marbury covers the cases for single document and multi-document text summarization. Moreover the property that a summary is produced for a particular user(s) and task(s) can be interpreted as if the definition also covers the query-guided and generic cases. The most important property that Mybury’s definition lacks is exclusivity. The same definition could also be applied to the task of keyword extraction. Although keyword extraction is occasionally also considered to be a text summarization task, but in general it is a distinct branch of text mining, as different from text summarization, the target document in keyword extraction is a collection of keywords and

34


not a text document consisting of coherent sentences. In his book Automatic Summarization, states Mani in 2001 that “a summary is a document containing several text units (words, terms, sentences or paragraphs) that are not present in the source document.”[10]. Consider a source document d and a target document t = d ∪ {s1 , . . . , sn } constructed by addition of n sentences to the source document. Clearly t is not a summary document and thus Mani’s definition lacks the universality property of a proper definition. The same problem also applies to the in the 2001 proposed definition of Sakai and SparkJones where they define a summary to be “a reductive transformation of a source text into a summary text by extraction or generation” [11]. The above definition, cannot be applied to the query-guided summaries and thus lacks the universality property of a definition. In 2002, Radev et al. defined a summary to be “a text that is produced from one or more texts, that conveys important information in the original text(s), and that is no longer than half of the original text(s) and usually significantly less than that”[12]. This definition lacks the generality property of a proper definition, as a restriction on the size of the output document is applied. An example of a recent attempt in defining automatic text summarization is the work of TorresMoreno in 2014 where he defines an automatic summary as “a text generated by a software, that is coherent and contains a significant amount of relevant information from the source text. Its compression rate τ is less than a third of the length of the original document”[13]. Torres-Moreno’s definition points to an important property of the summary text, which is its coherence. The definition does not concretize relevant information and is not applicable to query-quided summaries, resulting in a lack of universality property. It also lacks the generality property by introducing compression rate as a part of the definition. In Table 1 the properties of the discussed definitions are summarized. For the sake of readabil-

ISBN: 978-1-941968-26-0 ©2015 SDIWC

ity, following abbreviations are used in the table: U: Universality, G: Generality, M: Minimality, E: Exclusivity, R: Repeatability Table 1. Properties of Previous Definitions on Text Summarization U

G

M

E

R

Luhn [1]

7

X

X

7

7

Maybury [9]

7

X

X

7

7

Mani [10]

7

7

X

7

7

Sakai [11]

7

7

X

7

7

Radev et al. [12]

7

7

X

X

7

Torres-Moreno [13]

7

7

X

X

7

In the following section a commonly observed property of summaries, namely compression rate will be discussed and criticized as an unnecessary part of automatic summarization definition. 3

READABILITY

Usually in the context of automatic summarization we speak about compression rate τ which is defined as the ratio between the length of the summary and the length of the source document [13]: |summary| τ= , (1) |source| where |•| is the length of the document in characters, sentences or words. Various thresholds in the literature have been suggested for τ . In [14] a summary is defined to be a text which is not longer than the half of the source document. At the same time, in [15] the optimal compression ratio for a summary is defined to be between 15% and 30%. The use of compression ratio as an essential part of the definition has several disadvantages. The first one is that a direct comparison of the length of a summary and the length of the source document is not always possible. This comparison is commonly made based on the character, word, or sentence length. The choice between words, characters or sentences is normally made arbitrarily without any specific reasoning and typically depending on the underlying data set.

35


By using the count of sentences in source and summary document for the calculation of compression rate, the length of a sentence is completely ignored. So two sentences s1 and s2 with |s1 | |s2 | in the summary document will contribute the same amount to the computation of τ . On the other hand, computing the compression rate based on the length of documents in characters does not always return reliable information. As an example, consider two summaries t and t0 = t + ”.” where in t0 a punctuation mark is inserted at the end of t. According to the formula of compression rate, the compression rate of the first summary t is less than the compression rate of the second one t0 , although the summaries do not differ significantly and a human evaluator would not even notice the difference between the summary documents. Perhaps the most reliable measure among the introduced ones is the length of documents in words. However, to only consider the length of documents in word has its own drawbacks. One main drawback is that the complexity of the words will be completely ignored. This does not cause remarkable problems in the case of extractive summarization, but in the context of abstractive summarization where paraphrasing and lexical selection are typical procedures, this may make the measure inconsistent. From the other side, depending compression rate on the words (tokens), makes the comparison between the compression rates of different algorithms a tedious task. Different approaches use different techniques for tokenization of the underlying text and this results in the fact that compression rate will be highly dependent on the underlying tokenization algorithms. The second problem with the concept of compression rate is the need to define a specific threshold. This threshold is usually selected without any specific reasoning and mostly the selection is done in a way that the algorithm will return the best possible results for the underlying data set. As already discussed various thresholds are suggested in the literature and this arises the question whether specifying a hard-coded threshold is reasonable and con-

ISBN: 978-1-941968-26-0 ©2015 SDIWC

sistent with the natural way humans perform summarization? Alternatively, readability of a text seems to be a suitable substitution for the compression rate (note that in this work we do not use readability in its common sense meaning but rather we define it as a measure). In general, we expect from a summary to be more readable than the source document. The concept of readability can capture various dimensions such as the consumed time for reading the summary, its cohesiveness or the complexity of the vocabulary in it. The introduction of the readability may bring from one side more vagueness into the definition of the automatic text summarization, but from the other side, it will distract the focus from the compression rate that in the late researches was heavily regarded as a key factor, resulting in the ignorance of the other dimensions connected to the readability of a summary. Much research is already done to measure the readability of a text. This includes for example the Flesch Reading-Ease Score [16] that considers the average sentence length and the average word length in syllables: F RES =206.835 − 1.015 − 84.6

total words

total sentences total syllables total words

,

(2)

or the Gunning Fog Index [17] that considers the average sentence length and a list of hard words (words with more than two syllables) in the text: words GF I =0.4 sentences complex words + 100 . (3) words Of course a direct takeover of the existing readability scores in the field of automatic text summarization is not appropriate and more sophisticated multi-factorial scores considering reading time or cohesion of the summaries have to be designed. Some of the properties that should be required for a readable summary document are: • Time: The amount of consumed time for reading a summary document should be

36


less that the amount of time that has to be consumed for reading the original document. • Length: The length of a summary document should be less than the length of the original document (this can be modelled using the compression rate). • Cohesiveness: A summary document should be at least as cohesive as the original document. • Word Complexity: The average complexity of the words used in the summary text should be less than the average complexity of the word in the original document. Note that the proposed properties for a readable summary are a subset of all possible properties that a readable summary document can exhibit. The intention of introducing properties for readability is to emphasize that length of a summary document (as considered in compression rate) is only one of the crucial dimensions in text summarization and other dimensions (specifically amount of time consumed for reading the summary document in comparison to the original document) have to be considered too. Finally instead of compression ratio we suggest the use of readability ratio % = ρ(s) to compare ρ(t) the ease of reading of the source document ρ(s) to the summary document ρ(t). Beside the readability, content selection is another vital part of any definition on automatic text summarization. This will be discussed in more details in the following section. 4

CONTENT SELECTION

The decision which content to include in a summary is a critical one. The reason for this is that a summary will be finally read by a reader or a group of readers with diverse expectations from the content of the summary. Each reader has its own subjective preferences and expectations and creating a summary that fulfills all these subjective expectations is in practice impossible. With this is mind, summaries are commonly

ISBN: 978-1-941968-26-0 ©2015 SDIWC

classified into generic and query-guided summaries. Generic summaries are the ones that estimate user’s information need. In contrast guided summaries include the user’s information need and ignore other irrelevant parts of the source document [18]. The user queries form a set Q of concepts, aspects, keywords or entities formulated by the user, representing the user’s needs. Generic summaries shall be considered as a subcategory of query-guided summaries where the user query is an empty set Q = ∅. We claim that a generic summary is of lesser use and almost impossible to evaluate manually. Consider a document d with its corresponding summary t, where Q = ∅. Now consider two users U1 and U2 aiming to manually evaluate the quality of content selection in the summary t. At this stage users will answer the question whether the summary contains the most relevant (significant) information of the source document or not. This is exactly the place where the subjective preferences of U1 and U2 will be formulated on the fly. Thus for U1 we will have Q1 = {q1 , . . . , qn } and for U2 we will have 0 Q2 = {q10 , . . . , qm }. Having two different sets of queries (probably formulated unconsciously by the users) will make the comparison of the scores calculated for t by U1 and U2 inconsistent. By letting Q to be predefined, the vagueness of the phrases in the definition, such as a text containing significant amount of information, or a text containing important information in the original text will automatically disappear. The question is now how to handle the case Q = ∅? More specifically, how should phrases such as important or significant information be interpreted, although no user preferences are pre-defined? It is of course very restrictive to let the set of queries Q to be a non-empty set. This will lead to the lack of the universality property in the definition as the generic summaries can not be covered by such a restriction. By letting the set Q to be an empty set, the concretization of the keywords such as “important” or “significant” in the definition

37


will by postponed to the implementation of the algorithm itself. As an example, consider the work of Luhn [1]. For selecting the most important sentences in the document, Luhn followed an approach based on the frequencies of the words. Although Luhn’s approach is classified under the generic summaries, but a query such as q = {sentences containing most frequent words in document}, could also be used. Another example is the work of Edmundson et al. [19] that used the presence of cue words for content selection. This can also be formulated as a query set such as q = {”signif icant”, ”impossible”, ”hardly”}. In this way the evaluation of the summaries by multiple human evaluators will also be possible, as the evaluators will all be using the same set of queries. 5

DEFINITION OF AUTOMATIC TEXT SUMMARIZATION

Having told the drawbacks of the previous definitions of automatic text summarization and after discussing important aspects of any automatic text summarization system, we propose our own definition: Definition. Given a set Q of queries and a set K representing a knowledge base, automatic text summarization is a reductive transformation of a collection of documents D with |D|> 0 into a single or multiple target documents, where the target document(s) are more readable that the documents in D and contain the relevant information of D according to Q and K. The above definition exhibits the required properties of a proper definition as discussed in Section 1. The definition can be applied to indicative and informative summaries. Furthermore the set D is defined as a collection of source documents with cardinality greater than or equal to one which covers both single and multi-document summarization. The set Q consists of queries in the form of phrases, entities, sentences or keywords and it can also be an empty set. By this, both query-guided and generic summaries

ISBN: 978-1-941968-26-0 ©2015 SDIWC

are covered. This all results in that the proposed definition has the universality property. The introduction of a knowledge base K in the definition, covers the case of update summaries where the users need is only to get update information about a specific topic in a summarized manner. Note that K can also be an empty set leading to the case of generic and query-guided summaries. In the proposed definition it is also allowed to output multiple documents as the result of the summarization process. By this, the case of hierarchical summaries can be covered where the summary documents are ordered from more general and abstract ones, to more specific and detailed ones. In the proposed definition, no restriction is applied to the implementation of the algorithms. Summarization is defined as a reductive transformation, meaning that the target document should be always shorter than the source document and by the introduction of the concept of readability and elimination of the concept of compression rate, the generality property of the definition is guaranteed. The proposed definition also has the minimality property, as the elimination of any property in the definition will cause to the failure of the reconstruction of a text summary in a specific scenario. We claim that the proposed definition has also the exclusivity property as the definition is not applicable to relevant fields such as keyword extraction, natural language translation or topic detection. By use of the concept of readability, also the repeatability property of a proper definition is guaranteed. It is stated that the target document is more readable than the source document. Given a readable source document, it is always guaranteed that the target document is also readable. Assuming a source document consisting of two sentences and a target document consisting of one sentence produced from the source document, the question is now if the definition is still valid if we apply it to the target document consisting of one sentence? In other words, is it possible to summarize the target doc-

38


ument one more time, assuming the proposed definition? The requirement that the target document should always be more readable than the source document, results in the desired situation that some documents are not summarizable. This is especially the case where the source document is very short and compact enough that only a paraphrasing but not a summarization is possible.

6

CONCLUSION AND OUTLOOK

In this work we focused on a proper definition of automatic text summarization that has been neglected by many researchers in the community. We proposed the properties universality, generality, minimality, exclusivity, and repeatability (Section 1). Based on these properties, various existing definitions of automatic summarization in the literature have been investigated and criticized (Section 2). We also discussed important aspects of a proper definition of text summarization such as readability (Section 3) and content selection (Section 4). Finally in Section 5 we proposed our own definition of automatic text summarization and showed that the proposed definition exhibits all the properties of a proper definition introduced in Section 1. The proposed definition is by no means considered as a gold standard, however in the opinion of the authors it lacks many of the drawbacks of the previous definitions in the community. Many other features such as the language of a summary, its coherence or the way it has to be evaluated have not been discussed in this work as a part of a proper definition. A more detailed investigation of the existing work on automatic text summarization is needed to examine the need for a more complex definition of automatic summarization. Similar to any other definition, our proposed definition is also volatile in time and with respect to the community’s feedback.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

REFERENCES [1] H. P. Luhn, “The automatic creation of literature abstracts,” IBM J. Res. Dev., vol. 2, pp. 159–165, April 1958. [2] M. Kan, K. R. McKeown, and J. L. Klavans, “Applying natural language generation to indicative summarization,” in Proceedings of the 8th European Workshop on Natural Language Generation - Volume 8, EWNLG ’01, pp. 1–9, 2001. [3] D. D. and A. F. T. M., “A survey on automatic text summarization,” 2007. [4] N. Ketui, T. Theeramunkong, and C. Onsuwan, “An edu-based approach for thai multi-document summarization and its application,” ACM Trans. Asian Low-Resour. Lang. Inf. Process., vol. 14, pp. 4:1–4:26, January 2015. [5] J. Christensen, S. Soderland, G. Bansal, and Mausam, “Hierarchical summarization: Scaling up multi-document summarization,” in Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 902–912, June 2014. [6] Y. Meena and D. Gopalani, “Analysis of sentence scoring methods for extractive automatic text summarization,” in Proceedings of the 2014 International Conference on Information and Communication Technology for Competitive Strategies, ICTCS ’14, pp. 53:1–53:6, 2014. [7] S. Banerjee, P. Mitra, and K. Sugiyama, “Abstractive meeting summarization using dependency graph fusion,” in Proceedings of the 24th International Conference on World Wide Web Companion, WWW ’15 Companion, pp. 5–6, 2015. [8] R. McCreadie, C. Macdonald, and I. Ounis, “Incremental update summarization: Adaptive sentence selection based on prevalence and novelty,” in Proceedings of the 23rd ACM International Conference

39


on Conference on Information and Knowledge Management, CIKM ’14, pp. 301– 310, 2014.

multi-document summarization.,” Inf. Process. Manage., vol. 47, no. 2, pp. 227–237, 2011.

[9] M. T. Maybury, “Generating summaries from event data,” Inf. Process. Manage., vol. 31, pp. 735–751, September 1995.

[19] H. P. Edmundson, “New methods in automatic extracting,” J. ACM, vol. 16, pp. 264–285, April 1969.

[10] I. Mani, Automatic Summarization, vol. 3 of Natural Language Processing. John Benjamins Publishing Company, 2001. [11] T. Sakai and K. Sparck-Jones, “Generic summaries for indexing in information retrieval,” in Proceedings of the 24th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’01, pp. 190–198, 2001. [12] D. Radev and A. Winkel, “Multi document centroid-based text summarization,” in In ACL 2002, 2002. [13] J. M. Torres Moreno, Automatic Text Summarization. Wiley-ISTE, 2014. [14] E. Hovy and C. Lin, “Automated text summarization and the summarist system,” in Proceedings of a Workshop on Held at Baltimore, Maryland: October 13-15, 1998, TIPSTER ’98, pp. 197–214, 1998. [15] C. Lin, “Training a selection function for extraction,” in Proceedings of the Eighth International Conference on Information and Knowledge Management, CIKM ’99, pp. 55–62, 1999. [16] J. P. Kincaid, R. P. Fishburne, R. L. Rogers, and B. S. Chissom, Derivation of New Readability Formulas (Automated Readability Index, Fog Count and Flesch Reading Ease Formula) for Navy Enlisted Personnel. Research Branch report, Defense Technical Information Center, 1975. [17] R. Gunning, The Technique of Clear Writing. McGraw-Hill, 1952. [18] O. You, W. Li, S. Li, and Q. Lu, “Applying regression models to query-focused

ISBN: 978-1-941968-26-0 ©2015 SDIWC

40


Predictive Analytics of Student Graduation Using Logistic Regression and Decision Tree Algorithm Lagman, Ace and Ambat, Shaneth [email protected], [email protected] School of Computer Studies and Graduate Studies AMA University – Philippines

ABSTRACT Educational data mining is an emerging area of data mining application. It is concerned in describing and predicting patterns into huge amount of data usable to educational settings. One main topic in educational data mining is the student graduation. The student graduation rate is the percentage of a school’s firsttime, first-year undergraduate students who complete their program successfully. Almost half of first year freshmen enrolled in tertiary level failed to graduate. The colleges and universities consisting of high leaver rates go through loss of fees and potential alumni contributors. This study focused on two aspects; to compare the accuracy rate of different classification algorithms in predicting student graduation and to generate data models that could early predict, and to identify students who are prone of not having graduation on time. The results is use to design proper retention policies and help the student to graduate on-time.

KEYWORDS Data Mining, Decision Tree, Logistic Regression 1

INTRODUCTION

Knowledge and Data Discovery or KDD, is a field of computer science, which includes the tools and theories to help humans in extracting useful and previously unknown information from large collections of digitized data. One of the primary steps of knowledge discovery in databases is data mining. Data mining is used to discover patterns and relationships in data focused on large observational data bases. More recently, researchers and higher education institutions are also beginning to explore the potential of data mining in analyzing academic data. The goal of such endeavor is to find means to improve the services that these institutions provide and to enhance instruction. This type of

ISBN: 978-1-941968-26-0 ©2015 SDIWC

data mining application is more popularly known as Educational Data Mining or EDM. At present, EDM is more particularly focused on developing tools that can be used to discover patterns in academic data. This area of EDM is often referred to as Learning Analytics – at least as it is commonly compared to more prominent data mining approaches which process data from large repository for better decision-making [1]. However, EDM can also be used to improve the performance of HEIs. One example is in the area of enhancing student graduation. If EDM can be used to discern the patterns of attributes that lead to the increase/decrease of student graduation, administrators can use it to make sense of what is happening, and enable them to predict possible outcomes and take more appropriate action. According to Philippine Authority of Statistics, there is an imbalance between the students enrolment and students graduation. Almost half of the first full time freshmen students who began seeking a bachelor’s degree do not graduate. This scenario indicates the need to conduct research in this area in order to build models that can help improve the situation. This was validated by conducting initial research to a total of 1164 firsttime freshman students, enrolled from 2007 up to 2010 were taken as subjects in determining the graduation rate of this higher education institution. Currently, out from 1164 first time freshmen enrollees only 193 (16.6%) finished or graduate on time and 971 (83.4%) failed to graduate on time. Hence, the graduation rate of the first-time freshmen students in the institution under study validated the imbalance between student enrolment and graduation. Student graduation rate (SGR) is the percentage of a school’s first-time, first-year undergraduate

41


students who complete their program successfully. Studies show that most freshmen students enrolled in tertiary level do not graduate. Part of the reason is that, they are underprepared to make a successful transition from high school to college [2]. Reference [3], on the other hand, defines student retention as the “ability of a particular college or university to successfully graduate the students that initially enroll at that institution. Research studies from HEIs indicated that early identification of leaving students and intervention program are the key to understanding what factors lead to students graduation. Institutions should utilize Siedman’s retention formula for student success: Retention = Early (Identification) + (Early + Intensive + Continuous) Intervention. As such, early identification of potential leavers and successful intervention program(s) are the key for improving student graduation. Addressing this problem is critical because universities with high leaver rates go through loss of fees, tuition, and potential alumni contributors. The early identification of vulnerable students who are prone to drop their courses is crucial for the success of any retention strategy and helps improve and increase the chance in staying in course chosen. Reference [4] used predictive modeling for early identification of students at risk could be very beneficial in improving student graduation. Research shows that early identification of leaver students and intervention programs are key aspects that can lead to student graduation. The study aims to explore the utility of EDM in addressing the problem of student graduation for a HEI in the Philippines. This research thus encompasses the identification of at risk students that will enable administration to provide timely help. Data needed for the research to be processed will include student's pre-college data, demographic data, entrance examination and college data sets which include first year first term grades. 2

RELATED LITERATURE

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Data Mining is an application of a specific algorithm in order to extract patterns from data. It has become a very important process to convert this large wealth of data into business intelligence, as manual extraction of patterns has become seemingly impossible in the past. Data Mining is a step inside the KDD process, which deals with identifying patterns in data. In educational research, data mining has been used to study the factors leading students to choose to engage in behaviors which reduce their learning and to understand factors influencing university student retention and academic achievements. Logistic regression is a statistical method for analyzing a dataset in which there are one or more independent variables that determine an outcome. The outcome is measured with a dichotomous variable. It is a basic tool for modeling trend of a binary variable depending on one or several regressors (continuous or categorical). From the statistical point of view, it is the most commonly used special case of a generalized linear model. At the same time, it is also commonly used as a method for classification analysis [5]. Logistic regression is a predictive modeling technique that finds an association between the independent variables and the logarithm of the odds of a categorical response variable. This is one of the techniques used in analyzing a categorical dependent variable. It provides an association between the independent variables and the logarithm of the odds of a categorical response variable [6], [7]. However when a categorical variable has only two responses, then it is called binary logistic regression model. Decision tree learning is one of the most significant classifying techniques in data mining and has been applied in many areas, including business intelligence, healthcare, biomedicine, and so forth. The traditional approach to building a decision tree, designed by Creedy Search, loads a full set of data into memory and partitions, the data into a hierarchy of nodes and leaves. The two most popular algorithms are CART: Classification and Regression Tree (with generic versions often denoted C&RT) and Chi-Square

42


Automatic Interaction Detection or CHAID will be used to build the decision tree model. Decision trees are grown sequentially partitioning the entire dataset into sections using the method of recursive partitioning. The initial stage of the algorithm is called the split search. Data is partitioned according to the best split and this in turn creates a new second partition rule. The process goes on until there are no more splits. The resulting tree is known as a maximal tree. Decision tree builds classification or regression models in the form of a tree structure. It breaks down a dataset into smaller and smaller subsets while at the same time an associated decision tree is incrementally developed. The final result is a tree with decision nodes and leaf nodes. The topmost decision node in a tree which corresponds to the best predictor called root node. Decision trees can handle both categorical and numerical data. 3

METHODOLOGY

The researcher used the steps of Knowledge Discovery in Databases and CRISP-DM methodologies in creating the study. There are two-step processes of data classification. The first step is determining the training sets of data for training until a data model will be build that describes a predetermined set of classes or concepts. The second step is testing data to estimate the classification accuracy of the model. If the accuracy of the model is acceptable, the model can be used to classify future data instances for which the class label is not known. The researcher will be using logistic regression and decision tree in predicting student graduation 3.1 Data Sets and Attributes The data used in the study were enrollment, admission and grades data sets from school year 2007 up to 2010. The data was obtained from the office of Registrar, Admission and Management Information System of the university as shown in Table 1. The dataset includes pre-college and college data sets which will include only first time regular freshmen. Table 1. Attributes Description

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Variable Descriptions 1. Graduation status – Target Variable – Labeled 0 was coded for students who failed to graduate on time and 1 was coded for students who graduated on time. 2. Gender – Students Gender - Labeled 1 was coded for the male students and 2 was coded for female. 3. Location – Location of the Students – Labeled 1 was coded for students who are living in Metro Manila and 2 was coded for students who are living outside Metro Manila. 4. Scholarship - Financial assistance given by the school –Labeled 1 was coded for students who availed financial help and 2 was coded for students who were not given financial assistance. 5. Entrance Examination Results – The entrance examination were composed of Abstract, Verbal, Numeric and Science. The four categories of entrance examination were set as categorical particularly ordinal type of data sets. 6. First Year First Term Grade - The first year first term subjects were composed of Algebra, IT Fundamentals, Programming, English, Values Education and Physical Education. Values of this

43


section were set particularly ordinal.

as

categorical 3.3 Performance Measure of the Algorithm 3.3.1 Receiver Operating Characteristic (ROC)

3.2 Classification Algorithms 3.2.1 Logistics Regression Recoding was done on the graduation status variable such that 0=not graduated and 1=graduated. Recoding is important for easy interpretation of the logistic regression results). The recoding was done using SPSS v.21. Logistic regression analysis on all the hypothesized predictors were tested using six different methods which include forward conditional, forward left right, forward wald, backward conditional, backward left right, backward wald. A comparative of classification table will be shown per method used. These different methods generate a series of steps, where each step corresponds to a logistic regression model. A logistic regression model where all predictors which are significant (p 50 + 8m (where m = number of independent variables), but all samples are 25, so it is not possible to use this formula. In this study 25 experts and examiners were invited to participate. Finally, a total of 25 subjects were included which is the whole population.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

VI.

STUDY SETTING

Snow ball sample will be used because of small size of population and limit number of people interested in the study problem. Three departments that provide cyber crime fighting services are located in Amman. The self-administrated questionnaire will be used in this study: effectiveness of scientific methods in detecting phishing Websites Questionnaire. This questionnaire was adopted from adviser. It evaluates the effectiveness of knowledge and assessed attitude toward these methods used. Before starting data collection, the approval from Jordanian Public Security (PSD) was obtained. Prior to administration of the survey, participants received verbal and written information related to the nature and purposes of the study and were notified of their right not to participate, ability to withdraw at any time without penalty, and to not answer any questions as they wish. Furthermore, respondents were provided the chance to ask any questions at any stage. Also respondents were asked to read and sign a consent form prior to completion of the survey. Privacy and confidentiality were maintained as outlined in the consent letter. Knowledge and attitude toward scientific methods varied by their demographic characteristics and other factors were included in the questionnaire Inclusion criteria will be all Jordanian experts and examiner, have at least three years' experience in cybercrime field. Data was collected during the period of Dec 2012. VII. PILOT STUDY

A pilot study cannot be conducted because of small sample size which was appears during data collection for the three departments of PSD. A. method of Data Analysis The data will be analyzed using Statistical Package of Social Sciences (SPSS) for windows version 16. The quantitative data obtained from detective police and forensic laboratories investigators will be analyzed using descriptive statistics, including arithmetical averages to determine the response rate of individuals to questions from the questionnaire respondents to the study variables. Standard deviations will measure the degree of dispersion of the absolute values of the answers to the middle arithmetic and frequency for all answers. One Sample T- Test and One-Way ANOVA of the order to know whether there is a difference with statistical functions of the importance of items relating of variable A -value of less than 0.05 (a ≤ 0.05) will considered statistically significant. level low Medium high

1 - 2.49 Higher than 2.49 – 3.49 Higher than3.49-5

121


B. Proposed methods using URL (AUC)

The most effectiveness method used for reduction or preventing this problem

Our proposed method AUC, used C# oriented language to detect the phishing blacklists database of Phishtank organization for fighting phishing, by entering the URL of any site. It will check the URL comparing it with blacklists database to determine if it phish or not (Ham).

Can be via raising level of awareness and preconception of user.

This method has two advantages and one limitation: Advantage  The URL check used in the proposed method can determine the phishing Websites that may be classified as phishing sites.  It’s easy client side method classifying sites and fast given result, needs no training easy to use, easy installed and download in the browser.

The current law which concerns detecting and fighting phishing in Jordan is not considered a firm method and coordination between all concerning official site or private. For fighting and detecting phishing and others cybercrimes by the law are limited available. The study examines the effectiveness of scientific methods used in detecting phishing Websites by Jordanian experts and examiners at PSD The findings of the present study found that first-degree requirements need used new and updated methods besides the traditional one and the size of phishing problem in Jordan is so limited and recently under control but it cannot be neglected and could be increased in the future due to technology developments.

Limitations  The URLs of newly established phishing sites may not yet be included in the blacklist, and this can be avoided by updating the databases periodically. And this methods process of work as the following:  Open the AUC method and insert the suspect URL.  AUC will contact the phishtank organization database.  Then it will give the result of comparing with blacklist as phish or ham.

REFERENCES [1] [2] [3] [4] [5] [6] [7]

In our research 50 URL experimental samples were taken on 10-1-2013, 16:00 PM, 25 phish URL, 25 not phish

VIII. CONCLUSION Findings of the current study will be contributed to measure the effectiveness of scientific methods in detecting phishing Websites in Jordan. Additionally, also proposed new scientific method can be used in detecting phishing Websites depending upon URL to judge whether or not phishing or legitimate sites before entering that site based on blacklists of Phishtank’s direct link to the database. Furthermore, it sheds the light of access and source of methods and to initiate plans to modify these methods and attract experts and examiners of PSD attention to enhance the awareness to not fall in phishing by using different new methods to prevent or mitigate the phishing process.

[8]

[9] [10] [11] [12] [13] [14] [15] [16] [17]

The results will be a clear guidance and initiate plan to modify the methods used by enhance their ability to detect phishing.

[18]

The present method used is considered medium-level, depending on the sample point views, and they used spam filter methods in the first degree then password and finally toolbar. Particularly safe Google and other methods are considered very low percentage.

[20]

ISBN: 978-1-941968-26-0 ©2015 SDIWC

[19]

[21] [22] [23]

David Km. Website plagiarism and customer hijacking, 2005. Xi G. toward a Phish Free World: A Cascaded Learning Framework for Phish Detection. Jingguo W, Rui C, Tejaswini H, Raghav HR. An Exploration of the Design Features of Phishing Attacks. James SM, Macq, BA. G Dip Comp Deakin, 2011. Alnajim, AM. Fighting internet fraud: antiphishing effectiveness for phishing websites, detection. Doctoral thesis, Durham University 2009. Bergholz, A, Paaß, G, Reichartz, F, Strobel, S, Chang, J. Improved Phishing Detection using Model-Based Features. Wardman,B .A series of Methods for the Systematic Reduction of Phishing,2011. Chik, BW. Challenges to Criminal Law Making in the New Global Information Society: A Critical Comparative Study of the Adequacies of Computer-Related Criminal Legislation in the United States, the United Kingdom and Singapore, 2010. available at http://www.chariotsfire.com/thesis/Chapter2.pdf Wu M, Miller CR, Garfinkel L S. Do Security Toolbars Actually Prevent Phishing Attacks? ABA (American Bankers Association). Phishing Prevention and Resolution, July 2005 Reddy P V, Radha, V, Jindal,M. IJAEST 2011 ; v2, 39-45. The initiative for an open Arab Internet. Whittaker C, Ryner B, Nazif M .Large-Scale Automatic Classification of Phishing Pages. Sood. KS. Phishing Attacks: Challenge Ahead, 2012. Witte, N. Rating the Authenticity of Websites, 2011. White JS, Matthews N, Jeanna SL, John. A Method for the Automated Detection of Phishing Websites through both Site Characteristics and Image Analysis. Dhanalakshmi. R, Prabhu. C, Chellapan. C. Detection of Phishing Websites and Secure Transactions, IJCNS 2011; Volume-I, Issue-II. Likarish P, Dunbar D, Hansen ET, Hourcade PJ .B-APT: Bayesian AntiPhishing Toolbar, IEEE 2008. Kumaraguru, P,Sheng S, Acquisti A, Cranor FL, Hong J. Lessons from a Real World Evaluation of Anti-Phishing Training. Lagutin, D. Securing the Internet with Digital Signatures, 2010. Microsoft: Anti-Phishing Technologies Overview, 2007. Dhamija, R, Tygar, JD. Phish and HIPs: Human Interactive Proofs to Detect Phishing Attacks, 2005. 127-141

122


[24] Li, S, Schmitz R. A Novel Anti-Phishing Framework Based on Honeypots, IEEE 2009. [25] A Framework for Detection and Measurement of Phishing Attacks Researchers from John Hopkins University and Google Inc. [26] Ramachandran, A, Feamster, N, Krishnamurthy, B, Spatscheck, O Van der M, J. Fishing for Phishing from the Network Stream. [27] Sekaran R, Bougie U. Research Methods for Business. 2003; 170 (86) 4th Edition Wiley. [28] Almaiah, L. Majali: growth of crime in the Kingdom within the normal level by the Director of Public Security remarks AmmanNet ,2011. [29] Rousan, Basim. 1103 electronic crimes took place in Jordan in 11 months. Jordanian news 2011 . [30] Almaiah, L. Hackers. thieves by modern vision, 2010. [31] Alfakeir, M. Communications Minister Opens Information Systems Security Forum . today for development, Petra 2012. [32] Moukarzel, P. Academy Awards Arab Internet in the arabian region. Alghaad 2009. [33] Brigadier-Hamoud. Public Security race in the introduction of the electronic crime due to their importance and gravity. Salt News 2012 . [34] Nazal, E. 197 electronic crime since the beginning of the year and 35% of the perpetrators are non-Jordanians. Ammon 2009. [35] Tabachnic BG, Fidell LS .Tabachnick and Fidell formula. Using multivariate statistics.4th ed, New york:Harper Collins, 2001.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

123


Efficient Adaptive Tree-based Protocol for Wireless Sensor Networks Khaled Mousa Computer Engineering Department Pharos University in Alexandria [email protected]

ABSTRACT The usage of wireless sensor networks applications is becoming very important in monitoring and analyzing the environment by utilizing a number of sensor nodes. These sensing nodes are capable of collecting the required data and transmit it to the target application. The wireless sensor nodes use routing protocols for sending data between the nodes of the system as required by the application. These routing protocols vary in their effect on the node power consumption, the node life time and the system scalability. We propose an efficient adaptive tree-based wireless network routing protocol. The protocol uses selected nodes in each level of the network for routing to the next level. The new approach decreases the power consumption and increases the life time of the network in addition to improving the routing reliability.

KEYWORDS Wireless Sensor Networks; Routing protocol; Tree structure; Re-routing.

1. INTRODUCTION In this paper we propose a new network routing protocol. The proposed routing protocol is called Efficient Adaptive Tree-based Protocol (EATP). The aim of our routing protocol is to reduce the node power consumption and hence its life time. This is done by using few nodes in each level of the network for routing toward the sink node (the root of the tree). EATP is also reliable as it has a recovery procedure for any failed transmission path.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Amr Elsaadany Computer Engineering Department Pharos University in Alexandria [email protected] Basically, our routing protocol is query-based that allows alternative routes for the information to reach their destination. Each node will establish a main route to the Base Station (BS). During the communication with the base station, a node will use this recorded route and in case of communication failure, an alternative route will be used. The proposed protocol is based on the idea of a full tree that consists of a large number of nodes; each node has N children and one parent. The nearest node to the BS is the root of the tree and the BS is the parent of the root. The BS sends the query to the root, then the root retransmits it to its children, and each child node retransmits the query to its children and so on until the query is propagated in the whole network. If a node detects an event that matches the query, it sends a notification packet to its parent and the parent aggregates the data and retransmits the packet to its parent, and so on until the packet reaches the BS. Whenever a node receives a packet, it should send an acknowledgement (ACK) packet to the sender to assure receiving the packet. If a node tried to send a packet to its parent and did not receive an ACK packet, it tries again two more times and if no response from the parent, the sender considers the parent has died. In this case the sender needs to send the packet to the alternative parent node (all the way to the root). At this point, the tree is reconstructed. The rest of the paper is organized as follows. The next section provides the background of the common routing protocols. The section 3 explains the operation of the new routing protocol. Section 4 explains the protocol model and Section 5 discusses

124


the protocol evaluation. Section 6 includes the future work and concludes the paper.

2. ROUTING PROTOCOLS The existing routing protocols have many types and can be classified as follow: either based on functions, styles, or network structure.

2.1 According to their function mode The nodes can be classified [1] as proactive, reactive or hybrid. In proactive protocol the nodes continuously sense the environment and frequently transmit the data to the BS. LEACH [2] (Low Energy Adaptive Clustering hierarchy) is an example of this type of protocols. In reactive protocol the node transmit the data if there are sudden changes in the sensed value. TEEN [3] (Threshold sensitive Energy Efficient sensor Network) is an example of this case. In hybrid case the protocol can be proactive and/or reactive. APTEEN [2] (Adaptive Periodic TEEN) is an example of this type.

2.2 According to participating style of nodes The nodes can be classified as direct communication, flat or clustering protocols. In direct communication the node sends its data directly to the BS, and this consumes more energy. SPIN [4] (Sensor Protocols for Information via Negotiation) is an example of this protocol. In flat protocol if a node needs to transmit its data to the BS, it first searches for a valid route to the BS and then transmits the data. The Rumor protocol is an example in this case. In clustering protocols [5] [6], all the nodes are divided to number of clusters, each cluster has its own cluster head (CH), and this cluster head aggregates the data and communicate with the BS. LEACH is an example.

2.3 According to the network structure The nodes can be classified as hierarchical, data centric and location based protocol. Hierarchical

ISBN: 978-1-941968-26-0 ©2015 SDIWC

routing used to perform energy efficient routing, i.e., higher energy nodes can be used to process and send the information; lower energy nodes can be used for sensing the environment. LEACH, TEEN, APTEEN are examples of this type. Data centric [7] protocols are used to control redundancy which happens because sensor node does not have global identification number that can specify it uniquely. It depends on the naming of the desired data, thus eliminates the redundant transmission. SPIN and DD (Directed diffusion) are examples of this type. Location based routing protocols [8] need location information which can be collected from the GPS (Global Positioning System) to create an optimal path without using flooding technique. GEAR (Geographic and Energy-Aware Routing) is an example of this type. On the other had there are location-based routing protocols which are a family of routing protocols in which each sensor node should know its local location, for example, GPS coordinates. Also, each sensor node may know its remaining amount of energy. It uses these pieces of information during the forwarding of data packets from a source node to a destination node. This family of protocols includes GEAR and Min-Hop [9]. Geographic and Energy-Aware Routing (GEAR) is designed for routing queries to specific regions. All sensor nodes should be aware of their residual energy and their locations. Each node should know the location and remaining energy of its neighbours. GEAR uses the energy information to construct a heuristic function that avoids energy holes and choses sensors to route a packet to the target region without path failure. Minimum hop (Min-Hop) routing protocol forms an optimal path to send packets from a source node to the sink node. This optimal path is the route that has the shortest path to the sink. The path is represented by the number of hops to the sink. Therefore, the source node chooses the node of the next hop to be the node that has the minimum number of hops to the sink. If several nodes have the same number of hops to the sink, the node with the

125


highest residual energy is chosen. This process is repeated until the packet is reached to the sink node.

3. PROPOSED PROTOCOL In this section we describe the new protocol and list its characteristics. Figure 1 illustrates the network model.

Figure 1. Network Model.

The protocol utilizes a routing table at each node with the following characteristics:  Each node has a small routing table with several next hop addresses.  The network determines the routing table of each node during the setup phase.  The base station has a routing table that contains the addresses of all nodes and their coordinates. At the beginning, the network goes through a setup phase. The setup phase is consists of two steps.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

The BS starts first step similar to the PEQ (Periodic, Event-driven and Query-based) protocol [10] by sending a hop value to its nearest nodes. Each node updates its hop value, increment and retransmits it to its nearest nodes, and so on until a hop tree is constructed as shown in Figure 1 (a). The second step starts from the BS also by choosing the nearest neighbor to the BS to be the root of the full tree as shown in Figure 1 (b). The root chooses the nearest N nodes to be its children as shown in Figure 1 (c), then each child node chooses N nodes also to be its children and so on until the full tree is constructed as shown in the Figures 1 (d), 1 (e), 1 (f), but each chosen node as a child should be at the same hop level or at most in the next level of its parent to assure an efficient full tree. In case a node cannot find the nearest nodes in the same or in the next hop level, it will not have any children at all. Also each node should have an alternative parent in case of path failure. The alternative parent is determined when the node is selected as a child; by searching for the nearest node in the same hop level of its parent and chooses it as its alternative parent. The number of children N is a variable and its value assignment is depending on the shape of the location where the WSN network is applied. Once the setup phase has finished, the network is ready for the transmission phase. During this transmission phase the system in under operation; the queries are sent from the BS and the responses are sent back from the nodes. This phase consists of three components: 1. Query propagation. When the BS needs to detect a certain event in the environment, it sends a query packet to the root, the root retransmit the packet to its children, and each child retransmit the received packet to its children and so on until the query packet is propagated in the whole network. When a node receives a packet, it should send an Acknowledgement packet to the sender.

126


2. Event Delivery. When a node detects an event in the environment that matches the received query, it sends a notification packet that contains all the details of the detected event to its parent. The parent retransmits the received packet to its parent and so on until the packet is reached the BS. A certain situation may exist when two or more nodes have detected the same event because they are located in the same region (in close proximity). In this case each node that detected the same event will send a notification packet to its parent. As a result, there could be some duplicated packets going through the network causing unnecessary traffic. For this reason, all parent nodes aggregate the data received from their children and discards the duplicated packets. 3. Path recovery. If a node sent a packet to its parent and did not receive an ACK packet, it tries to send the packet again up to two more times. If the node did not receive an ACK again, it considers the parent node has died as shown in Figure 2. At this point, the node sends the existing packet to the nearest node in the next upper level on the way to the BS (and considers this node as its alternative parent node).

Figure 2. Path Recovery.

The alternative parent node will find that the received packet sender is not one of its children nodes, so the alternative parent will turn on the path failure flag and send the packet to its parent and so

ISBN: 978-1-941968-26-0 ©2015 SDIWC

on until the packet is reached the BS. The BS will find that the path failure flag is turned on, so it will restart the setup phase algorithm to reconstruct the tree to recover the failed path.

4. PROTOCOL MODEL In this section we describe the new protocol model. Basically, the BS node is defined as static with fixed coordinates. The total number of generated nodes and the number of children for each node are variables N and C respectively. The power consumption for sending a packet, receiving a packet and data aggregation are variables Sp, Rp, and Dag respectively. At the beginning, N nodes are generated with random coordinates and battery level. The setup phase starts by giving each node a hop value according to the distance between each node and the BS. For example, the nodes which have a distance between 1:3 meters give them a hop value of (1), the nodes which have a distance between 3:5 meters give them a hop value of (2) and so on. At this point, the hop tree is created and this is the first step of the setup phase. The second step starts from the BS by selecting the root node for the full tree which has the shortest distance to the BS. Then this root searches for the nearest C nodes to be its children, but these C children should be at the same hop level or at most in the next level. Afterwards, each of the chosen C children searches for the nearest C nodes to be their children and so on until the full tree is constructed. Now the setup phase is finished and we have a ready network for transmission phase. In the transmission phase the BS sends query packets to the root after a random time slot between each time. The packet will be propagated in the whole network, as each child that receives the query packet will retransmit it to its children and so on until the packet is fully propagated. In our model, a randomly chosen node will be considered as the event detector node. It will send a notification packet to its parent. The parent node will aggregate the data and retransmit it to its parent and so on until the packet is reached to the BS.

127


For each node that sends, receives or aggregates data, the model decreases its battery level by the corresponding value according to the assigned variables Sp, Rp, and Dag respectively. The model includes a procedure to check for the battery life of each node that needs to send, receive or aggregate the data. This procedure allows the node to perform the intended operation iff the residual battery level is still enough. If not, the procedure considers that the node has died. The packet under consideration is sent to the alternative parent (which turns on the path failure flag). As mentioned before, when the existing packet reaches the BS, the procedure which is responsible for reconstruction the tree will run to recover the failed path.

5. PROTOCOL EVALUATION  We have proposed and compared the new protocol with the other relevant protocols. We have found out that the new protocol (EATP) has the following features/advantages: 



EATP uses the energy of the nodes more wisely, as each node sends the packet to the nearest node and so there is no long-range transmission. This decreases the battery consumption and increases the battery life. Other protocols consume power in trying to transmit to the next node which could be at long distance from the sending node. As a matter of fact, in some protocols, nodes may have to transmit directly to the sink node. In the LEACH protocol, the some nodes consume more battery by sending data directly to the BS. Moreover, energy can be wasted due to the overhead of selecting the next cluster head nodes. EATP decreases the time delay to send packets, as each node has a very small routing table containing only the addresses of its children and its parent. So, no time needed to search for the shortest path or the

ISBN: 978-1-941968-26-0 ©2015 SDIWC





nearest node, as the nodes know already the next destination. Thus, our protocol saves the time (by not looking for the nearest node to transmit to) as opposed to the SPIN protocol which takes time in negotiation between the nodes to decide on the next hop before sending the actual data. In SPIN, the sender node sends the meta-data to the surrounding nodes as an advertisement. The interested neighbors can get the actual data by sending a request message. Finally, the advertiser node sends the actual data to the interested neighbors. This SPIN scenario does not happen in EATP. Also, the Energy Efficient Inter-cluster Communication-based protocol is another example of protocols that need time to search for the nearest node to send the packet between the clusters. EATP can guarantee the delivery of the packets by using the Acknowledgement mechanism (ACK). When a node sends a packet to a particular node(s), the sender turns on a timer and waits for an ACK. The receiver node should send a small ACK packet to the sender to assure receiving the data. The SPIN protocol does not guarantee the delivery of the data, which means that some nodes may miss the data. EATP has a high reliability as it has a path recovery algorithm to deliver the packets in case of path failure. The ACK mechanism is the key in detecting the path failure, as the path failure can be detected by not receiving the ACK packet. If the sender did not receive an ACK packet from the receiver it tries again two more times. If no ACK packet is received this means the receiver node has died and the path failure algorithm starts. This feature is not common in many protocols but we believe that it is a key feature of modern routing protocols. EATP solves one of the drawbacks in the PEQ protocol which causes unnecessary traffic. This traffic can be caused by not

128


aggregating the received data which allows for redundant packets in the network. In our protocol, each parent node aggregates the received data and discards the duplicated packets. This is extremely important as we know that the main design issues of sensors are the battery life and the delivery time. As it is clear from the above comparison, EATP can guarantee the packet delivery while maintaining the energy level of the nodes by the fact that it reduces the transmission distance and processing time due to fast routing decision by each node.

6. CONCLUSION The proposed routing protocol is named Efficient Adaptive Tree-based Protocol (EATP). This protocol uses the energy of the nodes more efficiently, as each node sends the packets to its nearest node(s). This means that there is no long-range transmission and this increases the battery life. The protocol also decreases the delay time to send the packets, as each node uses a small routing table (that has been built during the setup phase). This routing table contains the addresses of the nodes' children and the parent and so, no time needed to search for the shortest path to the BS or the nearest node. The new protocol has high reliability as it has a path recovery algorithm to deliver the packets in case of path failure by sending them to the alternative parent node. When this happens, the BS reconstructs the tree to recover from this exceptional case of failed path. Also, each parent node performs an important function to decrease the network traffic, as the parent nodes aggregates the received data and discard the duplicated packets to prevent any duplicated packets from being sent multiple times to the BS. As future work, we plan to add the following capabilities: (1) a function to propagate variety of the queries in the whole network, (2) a function to query or update the battery level of each operating node, (3) a function to collect statistics about the overall status of the system.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

In conclusion, we propose the usage of our protocol in many applications that can be modeled as tree structured systems. This is quite possible for new systems being implemented in new scalable industries that can take advantages of the features provided in our new protocol.

REFERENCES [1] S. Pal, D. Bhattacharyya, G. S. Tomar and T. Kim, "Wireless Sensor Networks and its Routing Protocols: A comparative Study," IEEE Computer society, pp. 314-319, 2010.

[2] S. Naeimi, H. Ghafghazi, C. Chow and H. Ishii, "A Survey on the Taxonomy of Cluster-Based Routing Protocols," MDPI, pp. 7350749, 2012.

[3] D.P. Agrawal and A. Manjeswar, "TEEN: A protocol for enhanced efficiency in wireless sensor networks," 1st International Workshop on parallel and distributed Computing issues in wireless networks and mbile computing, pp. p-189, 2001.

[4] L. J. G. Villalba, A. L. S. Orozco, A. T. Cabrera and C. J. Barenco, "Routing Protocols in Wireless Sensor Networks," MDPI, pp. 8399-8421, 2009.

[5] A. Martirosyan, A. Boukerche and R. W. Nelem, "A taxonomy of Cluster-based Routing Protocols for Wireless Sensor Networks," IEEE Computer Society, pp. 247-253, 2008.

[6] X. Liu, "A Survey on Clustering Routing Protocols in," MDPI, pp. 11113-11153, 2012.

[7] D. Geetika Dhand, "Survey on Data-Centric protocols of WSN," IJAIEM, pp. 279-284, 2013.

[8] S. Y. Agrawal and P. C. M. Raut, "A Survey on Location Based Routing Protocols for Wireless Sensor Network," International Journal of Emerging Technology and Advanced Engineering , pp. 123-126, 2013.

[9] A. M. El-Semary and M. M. Abdel-Azim, "New Trends in Secure Routing Protocols for Wireless Sensor Networks",International Journal of Distributed Sensor Networks, Article ID 802526, 16 pages

[10] A. Boukerche, R. W. N. Pazzi and R. B. Araujo, "A Fast and Reliable Protocol for Wireless Sensor Networks," Canada Research Chair Program, NSERC, Canada Foundation for Innovation Funds and OIT/Distinguished Researcher Award, pp. 157-164, 2004.

129


Voice Communications over WLAN on Campus: A Survey Munira A. AlShebel, Amal S. Alamr, and Amjaad H. Alshammari Imam Muhammad ibn Saud Islamic University College of Computer Science and Information

[email protected], [email protected], [email protected]

ABSTRACT

1 INTRODUCTION

Voice over Wireless Local Area Network (VoWLAN) technology offers significant benefits that provide wireless voice applications for universities, hospitals and enterprises. Some of these primary benefits are cost efficiency, noticeable increase in productivity rate and wirelessly accessing employees anywhere. There are two technologies that VoWLAN consist of, first is the Wireless Local Area Network (WLAN), while the second is the Voice over Internet Protocol (VoIP). Although, VoWLAN requires more than just a WLAN; it is in need of an Internet Service Provider (ISP) to access the Internet [1, 2]. Voice call communication over Wi-Fi is another network technology that provides free calls over WLAN without the need of an ISP [3]. In this paper, we present a survey concerning voice communications over various network technologies and suggest a solution as a result of comparison between these technologies. This solution aims to provide a facility of making free voice calls that would only use WLAN technology without the need of a service provider. Using free bands of WLAN would provide no cost internal communications and would include the benefits of VoWLAN. The challenges that face deployment of this solution would be providing mobility and large coverage area. Therefore, WLAN will need to be extended by including several Access Points (APs).

Enterprises, hospitals and universities have adopted WLAN technology quickly, freeing them from the cost of wiring a building for voice calls or data traffic, also improving members' productivity by allowing mobility [4]. VoWLAN is a new technology that combines two popular technologies: WLAN and VoIP. Meaning, VoWLAN systems are an extension of wired VoIP systems and they offer significant benefits of providing mobility, lower cost and the liberty to access applications wirelessly. Researches, however, show that most wireless usage of voice calls and data traffic take place indoors [2]. Moreover, in universities, most of the communications that take place are either between instructors and students or instructors internally.

KEYWORDS

To further elaborate, 802.11 can operate on both infrastructure and ad hoc modes. Ad hoc mode is a decentralized communication between wireless devices; consequently, there is no need for an AP [7, 5]. Presented in Figure 1 are the components of the infrastructure mode and its parts as mentioned in 802.11 [7]. A number of nodes, called stations (STAs) are

Network technology, free voice call, WLAN, Wi-Fi, voice message, VoIP, VoWLAN, AP, QoS, 802.11, ESS.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

802.11 is a wireless standard that requires no license since it uses free portions of the radio spectrum [5]. Wireless Fidelity (Wi-Fi) is a certified version of the 802.11 standard, this means that Wi-Fi is the name of 802.11 standard variants that ensures interoperability between devices. Wi-Fi is important; it ensures compatibility and reliability for all Wi-Fi certified devices [6].

130


connected to an AP. The STAs and the AP that are in the same coverage form a Basic Service Set (BSS). A group of BSSs that are connected via a distribution system is called an Extended Service Set (ESS) [4].

There are some challenges that affect voice transmission over various network technologies and design issues to ensure campus coverage when using a WLAN. 2.1 Voice Characteristics and Requirements

Researches on voice call communication over Wi-Fi use a specific type of WLAN with 802.11 standard. Unlike many other wireless standards, 802.11 runs on free portions of the wireless radio band, this means that no license is required to communicate using 802.11 WLAN [3]. Currently, millions of people throughout the world are using 802.11 standard, to wirelessly connect to the Internet when they are on the road. They are also using 802.11 standard to create private wireless networks in their homes and offices [6]. An example of the former using is VoWLAN that provides cost saving by 802.11 WLAN. However, VoWLAN requires payment for an ISP in order to access the Internet [2]. Voice call communication over Wi-Fi, on the other hand, is an example of the latter one. Using unlicensed bands of 802.11 WLAN without the need of a service provider will ensure “free” internal communications [3].

Voice traffic is different from data traffic. This section enumerates the characteristics of voice traffic: low bandwidth, delay sensitivity, small packet size and sporadic nature [8]. 2.1.1 Low bandwidth: Voice transmission does not need much bandwidth. 802.11 gives all competing devices equal probability to access the medium. As a result, data traffic will always gain ample bandwidth compared to voice traffic. 2.1.2 Delay sensitivity: Data traffic is relatively delay insensitive. In contrast, voice traffic can only tolerate end-toend delays of 200 to 400 ms. Furthermore, voice traffic is sensitive to delay jitter between packets. 2.1.3 Small packet size: Packets of voice traffic should be small, up to tens of bytes. Each packet has a notable number of metadata. 2.1.4 Sporadic nature: In voice communications, one of the two channels (uplink and downlink) is usually idle and each channel is used in a short period of time called Talk Spurt.

Figure 1. Architecture of the Infrastructure Mode

2 VOICE TRANSMISSION CHALLENGES AND DESIGN ISSUES

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Voice transmission requires a high quality of communication such as the high speed and the low level of interference. Fortunately, while 802.11 standard rises, it is improving and becoming faster, more secure, and will meet the voice transmission requirements [6].

131


The requirements of campus coverage comprise: using multiple APs and channel reuse [9].

Voice communications over network technologies can be divided as follows: VoIP, VoWLAN, voice transmission over LAN using Bluetooth and voice call communication over Wi-Fi.

2.2.1 Using multiple APs:

3.1 VoIP

One of the WLAN goals is allowing users to connect to that WLAN from any area covered by an AP. Therefore, WLAN contains a handful of APs. When APs are connected together in an ESS, users can roam the coverage area of an AP to another AP without needing to reauthenticate. In order to achieve this, all of the APs in an ESS must be able to communicate and transfer information to each other.

In [10], Mehta et al. introduce the concept of VoIP technology and identify its benefits. They describe VoIP components, types of voice coders and reliable mechanisms. They also enumerate some voice transmission issues as: Quality of Service (QoS), packet loss, jitter and latency. Finally, they give a brief comparison between H.323 and Session Initiation Protocol (SIP) as signaling protocols.

2.2.2 Channel reuse:

This reference is a theoretical paper that introduces VoIP components and mechanisms. The paper considers that packet switching networks are more suitable than circuit switching for voice transmission but with clever implementations. VoIP has a lot of chances to improve, for instance, it could enhance QoS, use reliable mechanisms and reduce the running cost.

2.2 Campus Coverage Requirements

Figure 2 illustrates the scheme that is used for channel reuse [9]. As it mentioned, to cover a large building, many APs will be used. The aim of the design is to distribute APs in such a way that user's station will be connected to at least one AP. There are eleven available channels, to avoid interference between APs, APs are separately adjacent by three to six channels. For a multistory building, consider interference from APs on the floors above and below [9].

Figure 2. Channel Reuse Scheme

3 VOICE COMMUNICATIONS OVER NETWORK TECHNOLOGIES

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Architecture for VoIP model is outlined in [11]. The authors evaluate different architecture models based on VoIP services. VoIP services include: management services, basic VoIP services and supplementary services. They focus on compatibility and independence between VoIP providers. This reference is an analytical paper that uses prototypes to describe and evaluate different architecture models. The paper introduces some factors that affect using VoIP applications. These factors include: updating VoIP software is costly and affects both clients and servers, adding a new service needs installation and configuration, compatibility between clients is not always achievable and the cost of installation and maintenance of VoIP would probably be unacceptable for small companies.

132


In [12], the authors describe the design of a VoIP project. They divide the document into three parts. First part includes: VoIP module, SIP, Reliable Transfer Protocol (RTP) and RTP Control Protocol, while the second part describes system software design and the third part gives a brief description on the hardware design.

properties. Server establishes and maintains a session for each connection.

This reference introduces the design as three layers: application layer, system layer and hardware layer. They choose design elements for each layer carefully, however, this project needs an ISP to operate.

High-Quality Voice over WLAN is a project developed by Cisco in [15]. This project discusses design issues to deploy WLAN infrastructure for supporting voice communications. It provides solutions for highquality voice calls such as using wireless "overthe-air" QoS. It also provides different roaming mechanisms.

In [13], Luiz et al. compare the performance of different applications for VoIP. They use broadband, connection forms with the receiver, delay and the signal quality as parameters for comparison. They analyze four applications: Google talk, MSN Messenger, Skype and Yahoo Messages. This reference is an empirical paper that describes in details some VoIP applications. VoIP enables two remote devices to communicate over the Internet. The application of an IP device converts voice signals to be transmitted on data networks. Even though the communications that take place via IP devices over the data network exist, these IP devices are fixed. Thus, they lack in users' mobility. Also, these devices require payment for an ISP to access the Internet. The paper shows that Skype is the best application for VoIP since it guarantees the quality of voice with least data loss. 3.2 VoWLAN Ravindra et al. [14] produce a better solution to transmit voice by using WLAN to access the Internet. Installation and maintenance of a WLAN connection are much easier and more cost effective than a wired LAN. They use the client server model to accomplish system's

ISBN: 978-1-941968-26-0 ©2015 SDIWC

This reference is a theoretical paper that provides flexibility and mobility to voice communications' users by using WLAN, although, VoWLAN depends on an ISP to work.

This reference is a project that deploys WLAN to provide VoIP with high quality as in wired networks and support mobility and flexibility of WLAN. However, the project needs to use an ISP which increases the running cost. In [16], the authors evaluate VoWLAN on different IEEE 802.11 standards. First, they develop VoIP system using open source applications and plentiful clients to provide voice traffic. Next, they implement the system on 802.11 b/g/n. Then, they use delay, jitter and packet loss as parameters to compare different voice qualities. Finally, the voice qualities are figured into graphs. This reference is an empirical paper that analyzes quality of VoIP using different 802.11 standards. It also shows several factors that affect the voice quality such as the wireless standard, the environment, the difference between a VoWLAN device and the AP and the interference from other devices. Google Hangouts [17] provides voice calls using VoWLAN technology. VoWLAN becomes the main communications channel between small companies' employees.

133


VoWLAN applications, such as Google Hangouts, support users' mobility and large area coverage. These applications have some limitations, such as the number of contacts per identity, also, to gain their services, the application has to access to the Internet connection. In [18], the authors test the performance of VoWLAN using Skype, X-Lite and WirelessMon for wireless monitoring. VoIP applications prove to be desirable, like Skype and Viber and many other applications, because of its low cost, as well as VoWLAN that requires no cost and easy to implement. This reference is an empirical paper that compares the behavior of VoWLAN against VoIP over a wired connection and as a result, the authors found that the voice quality is identical for the both of them, but with a longer delay on VoWLAN. The paper shows that the performance of the hand-off between one AP to another takes one second. The authors found that the connection and the voice quality were satisfying when in a very large area. Additionally, they found that the connection is stable unless the power is decreased. 3.3 Voice Transmission over LAN Using Bluetooth Voice Communications using Bluetooth Project proposed in [19] uses 2.4GHz free band. Bluetooth is used as a medium for voice communications where two cell phones, connected via Bluetooth, can make a voice call. Bluetooth covers a very short distance of approximately ten meters. The project provides free calls, but a short range of mobility. Therefore, it limits users' mobility and its communication is far slower than WLAN. 3.4 Voice Call Communication over Wi-Fi

possibilities of these ways. In VoIP, each IP phone needs an ISP, which is costly to use. Voice over traditional intercom system discards the mobility of the user. Voice over Bluetooth has a range of up to one meter or three feet, which is a very limited. The project's aim is producing free call communications with various features. Therefore, they develop an Android application that offers free voice calls over Wi-Fi system within a WLAN. They establish four types of modules: server, sender, receiver, and router that acts as an AP. This reference is an empirical paper that enhances the use of WLAN, by reusing the existing Wi-Fi networks. WLAN is cheaper, faster, and easier to setup. It provides free voice calls and free voice messages to both available and unavailable users with no need for an ISP, hence it is free. However, this project uses only one router. As a result, it limits the users' mobility. The ability of installation and maintenance in a wireless LAN is easier than a wired LAN. The proposed project in [1] generates a personal computer (PC) application with a free voice call, video call and chatting using a wireless LAN. SIP is used for voice transmission and Java Media Framework (JMF) package is used for video transmission. A server with one gigabyte Random Access Memory (RAM) is responsible for distributing connections between the clients. This reference is an empirical paper that adds an authentication service and encryption and decryption algorithms to enhance the security. Chatting history for each client is saved and the client can send messages to offline clients. However, the proposed project eliminates the mobility of users. To provide a facility of free voice communications between end users in an

Implementation of voice communications has various ways; authors in [3] considered the

ISBN: 978-1-941968-26-0 ©2015 SDIWC

134


Table 1. Classification and Comparison of Various Network Technologies Technology

References

VoIP

[10], [11], [12], [13]

VoWLAN

Mobility

Message storage

Large area coverage

No

No

Yes

Yes

[14], [15], [16], [17], [18]

No

High

Yes

Yes

Voice Transmission over LAN Using Bluetooth

[19]

Yes

Low

No

No

Voice Call Communication over Wi-Fi

[3], [1], [20], [21]

Yes

Low

Yes

No

organization and secure communications from outsider influence, the proposed project in [20] develops a wireless LAN to enable Wi-Fi devices to communicate through 2.4GHz free channel. If the destination user is out of the network range, the system proposes converting the calls to Global System for Mobile Communications (GSM) service provider; it depends on the Subscriber Identification Modules (SIM) card for the user. The project produces free calls, secures the caller from insiders hacking and offers shortrange mobility. If a sender or a receiver goes out of Wi-Fi range, the call is tunneled through GSM service provider and it will not be free. The proposed system only covers a small area. Before establishing a voice communication, the user has to choose whether to make the call over GSM service provider or through Wi-Fi. The suitable way to transmit data and voice over a network at no cost is through WiFi wireless LAN. In [21], the project offers three ways of voice communications: either over Wi-Fi Peer to Peer (P2P) connection within the sender’s Wi-Fi region, or virtual connection through an AP based on 10 mobile digit numbers or across the GSM server

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Free voice calls

provider. The system converts the mobile phone number to a unique Internet Protocol (IP) address using IPv6. This reference is an empirical paper that avoids IP collision in communications over WiFi and provides a cross-platform. Furthermore, it turns laptops/PCs into a Wi-Fi wireless AP to cover a large area and support mobility. However, it does not offer voice massages recording and when using a laptop as an AP it will consume power. 4 COMPARISONS SUMMARY Table 1 shows how different references are categorized under various network technologies and also summarizes the comparison between these technologies according to their different features. Note that there is no technology that provides all four features: free voice calls over an unlicensed band, mobility, message storage for unreachable users and large area coverage. Therefore, we suggest using WLAN, as a medium for communications within the campus, as there is no need for an Internet access for these communications. In this manner, we would eliminate the cost of VoWLAN. Also, using only one network

135


technology, WLAN, would guarantee higher QoS, such as low delay, high throughput and low packet loss, than the technologies that are using the Internet. However, the challenges for this solution are providing mobility and campus coverage. WLAN will be extended to perform call handoffs between multiple Wi-Fi routers (APs). The proposed solution will operate 802.11 on infrastructure mode and use ESS to ensure the coverage of the area on campus. Therefore, it will cover the entire campus and increase the mobility. Moreover, this extending WLAN will provide voice storage messages and text messages as voice calls, with no need of an ISP.

[5]

[6]

[7] [8]

[9] [10] [11]

[12]

5 CONCLUSION [13]

In this paper, we present a survey regarding voice communications over various network technologies and summarize the comparisons between these technologies according to various parameters. We also discuss the benefits of using WLAN for communications on campus. Finally, we provide a solution that enhances the use of WLAN as a medium for voice transmission instead of a medium for Internet access, which would support free voice calls, messages, mobility and large area coverage.

[14]

[15]

[16]

[17]

[18]

REFERENCES [19]

Akshay Iyer, Akshay Badgujar, Maheshkumar Eaga, and Rohit Iyer, “Voice and Video over Wireless LAN”, International Journal of Scientific and Research Publications, Vol. 3, issue. 9, September 2013, pp. 1 – 5. [2] Tech Target Network, “The role of VoWLAN solutions”, http://searchunifiedcommunications.techtarget.com/f eature/The-role-of-VoWLAN-solutions, October 2008. [3] Omkar Manjare, Sagar Bamnikar, Prathamesh Deshmane, Om Dongre, and Dr. Preeti Patil, “Voice Call Communication Over Wi-Fi”, International Journal of Engineering Research & Technology, Vol. 2, issue. 7, July-2013, pp. 2259 – 2263. [4] Frank Ohrtman. Voice Over 802.11. London: Artech House, 2004, 6. [1]

ISBN: 978-1-941968-26-0 ©2015 SDIWC

[20]

[21]

Mohammad Ilyas, and Syed Ahson. Handbook of Wireless Local Area Networks. USA: CRC Press, 2005, 77-87. Harold Davis. Absolute Beginner's Guide to Wi-Fi Wireless Networking, USA: Que Publishing, 2004, 7-22. Jochen Schiller, Mobile Communications, Second edition, Addison Wesley, 2003. Usman Ismail. “Virtual PCF: Improving VoIP over WLAN performance with legacy clients,” A thesis presented to the University of Waterloo in fulllment of the thesis requirement for the degree of Master of Mathematics in Computer Science, Waterloo, Ontario, Canada, 2009 James Trulove. Build Your Own Wireless LAN. USA: McGraw-Hill, 2002, pp. 242-277 Princy Mehta, and Sanjay Udani, “Voice over IP,” IEEE, 2001, pp. 36-40 Markus Hillenbrand, Joachim Götze, and Paul Müller, “Voice over IP – Considerations for a Next Generation Architecture,” IEEE, Germany, 2005. Sarfraz Nawaz, Mark Niebur, Scott Schuff, and Athar Shiraz, “VOIP Project: System Design,” Columbia University, Columbia, 2009. Luiz Filho, Maircio Costa, and Rogenio Filho, “Performance and Quality of Service on Free Softwares for VoIP,” The 1st International Conference on Next Generation Network, Korea, 2006. Ravindra Pardhi, and Vishwas Gaikwad, “Module wise Design of Voice and Video over Wireless LAN,” International Journal of Computer Trends and Technology, India, 2011. Cisco Systems, Inc., “Design Principles for Voice over WLAN: A White Paper,” USA, Cisco Public Information. Nor Ibrahim, Mohd AbdRazak, AbdulHalim Ali, and Wan Yatim, “The Performance of VoIP Over IEEE 802.11,” Universiti Kuala Lumpur British Malaysian Institute, Malaysia, 2013. Hangouts.google.com, 'Google Hangouts', 2015. [Online]. Available: https://hangouts.google.com/. [Accessed: 14- Nov- 2015]. Lozano Gendreau, Antoun Halabi, Maya Choueiri, and Valery Besong, “VoWF (Vo-IP over Wi-Fi),” IEEE, 2006. Arun Biradar, Dr. Ravindra Thool, and Dr. Rajappa Velur. “Voice Transmission over Local Area Network Using Bluetooth,” IEEE, 2009, pp. 1-6 Venkatraman, Siddharth Natarajan, and Padmavathi. “Voice Calls Over Wi-Fi,” presented at Proceedings of the World Congress on Engineering and Computer Science, San Francisco, USA. 2009. Shyam Sundar, Krishna Kumar, Selvinpremkumar and CHINNADURAI (Ph.D Scholar), “Voice over IP via BLUETOOTH/WI-FI Peer to Peer,” IEEE, 2012, pp. 828-837.

136


A 3-Dimensional Object Recognition Method Using Relationship of Distances and Angles in Feature Points Seiichi Maehara, Kazuo Ikeshiro and Hiroki Imamura Department of Information Systems Science, Graduate School of Engineering, Soka University Mailing Address: 1-236, Tangi-machi, Hachioji-Shi, Tokyo, Japan 192-8577 E-mail: {e14m5225, e10d5201}@soka-u.jp, [email protected]

ABSTRACT In recent years, human support robots have been receiving attention. Especially, objects recognition task is important in case that people request the robots to transport and rearrange an object. We consider that there are five necessary properties to recognize in domestic environment as follows.  Robustness against occlusion  Fast recognition  Pose estimation with high accuracy  Coping with erroneous correspondences  Recognizing objects in different aspect ratio As conventional object recognition methods using 3-dimensional information, there are modelbased recognition methods such as the SHOT and the Spin Image. The SHOT and the Spin Image do not satisfy all five properties for the robots. Therefore, to satisfy the five properties of recognition, we propose a 3-dimensional object recognition method by using relationships of distances and angles in feature points. As our approaches, firstly, the proposed method uses a curvature as a feature in a local region. Secondly, the proposed method uses points having high curvature as feature points. Finally, the proposed method generates a list by listing relationship of distances and angles between feature points and matches lists.

KEYWORDS 3-dimensional object recognition, Feature points, Relationship, Curvature, List

1 INTRODUCTION In recent years, human support robots have been receiving attention [1]-[3]. Then, the

ISBN: 978-1-941968-26-0 ©2015 SDIWC

robots are required to perform various tasks to support the human. Especially, objects recognition task is important in case that people request the robots to transport and rearrange an object. We consider that there are five necessary properties to recognize in domestic environment as follows.  Robustness against occlusion  Fast recognition  Pose estimation with high accuracy  Coping with erroneous correspondences  Recognizing objects in different aspect ratio Firstly, the robots need the robust recognition for occlusion because occlusion occurs between different objects in domestic environment. Secondly, the robots need to recognize a target object fast to achieve required tasks fast. Thirdly, the robots need to estimate a pose of a target object with high accuracy to manipulate a target object. Fourthly, the robots need to cope with erroneous correspondences to recognize objects which have the same feature in a local region but which are not the same object as shown in figure 1. Finally, the robots need to recognize objects which are similar shape but which have the different aspect ratio such as a cube and a rectangular parallelepiped. As conventional object recognition methods using 3-dimensional information, there are model-based recognition methods such as the SHOT [4]-[5] and the Spin Image [6]-[8]. Table 1 shows properties of these methods. These methods have robustness against occlusion and are able to estimate a pose of an object with high accuracy even if occlusion occurs on

137


Table 1. Properties of conventional methods and the proposed method.

aspect ratio because different aspect ratio are distinguished by distances and angles which are registered as relationships between feature points. Figure 1. An example of objects which have the same feature but which are not the same object.

2 PROPOSED METHOD

objects by using features in a local region. Especially, the SHOT is able to recognize objects fast by using sparse feature points. However, the SHOT and the Spin Image misrecognize objects which have the same feature in a local region but which are not the same object because these methods do not cope with erroneous correspondences as shown in figure 1. In addition, these methods misrecognize objects which have the different aspect ratio. As mentioned above, as shown table 1, the SHOT and the Spin Image do not satisfy all five properties for the robots. Therefore, to satisfy the five properties of recognition, we propose a 3-dimensional object recognition method by using relationships of distances and angles in feature points. As our approaches, firstly, the proposed method uses a curvature as a feature in a local region to have the robustness against occlusion. Secondly, the proposed method uses points having high curvature as feature points to recognize objects fast. Finally, the proposed method generates a list by listing relationships of distances and angles between feature points and matches lists. Thereby, the proposed method estimates a pose of a target object with high accuracy and copes with erroneous correspondences by using not only the feature points but also relationships between feature points. Furthermore, the proposed method recognizes objects in different

2.1 Flow of the Proposed Method

ISBN: 978-1-941968-26-0 ©2015 SDIWC

In this section, we describe about an overview of the proposed method based on its processing flow. As shown figure 2, the proposed method consists of a list generating process of a target object and a recognition process of the target object in a scene. In the list generating process (Figure 2 (a)), the proposed method firstly registers 3dimensional data of the target object as a teaching data. Secondly, the proposed method extracts feature points based on curvature from the teaching data. Finally, the proposed method generates a list by listing relationship between each extracted feature points. In the recognition process (Figure 2 (b)), the proposed method firstly extracts candidate regions from an inputted scene data. Secondly, the proposed method extracts feature points and generates a list from the candidate region in the same way as the list generating process (Figure 2 (a)). Thirdly, the proposed method associates the feature points of the teaching data with the feature points of the candidate region by matching their lists. Finally, the proposed method recognizes the target object and estimates a pose of the target object in the scene data by applying a rigid registration using associate feature points. We represent each process in the next sections.

138


Start Input teaching data Feature points extraction List generating

End (a) List generating process of a target object.

Start Input scene data Candidate regions extraction Feature points extraction List generating List matching Rigid registration

End (b) Recognition process of a target object. Figure 2. The flow of the proposed method.

(a) 3-dimentional (b) The result of (c) The result data of a target extracting feature of extracting points having a feature points. object. curvature more than 𝒕𝒉. Figure 3. Illustration of extracting feature points.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

2.2 Feature Points Extraction Firstly, the proposed method calculates a curvature 𝐾 in each point of the teaching data by using 𝜆0 (1) 𝐾 = , 𝜆0 + 𝜆1 + 𝜆2 𝑁

1 ̅)(𝒑𝑖 − 𝒑 ̅ )𝑇 . 𝑪 = ∑(𝒑𝑖 − 𝒑 𝑁

(2)

𝑖=0

In equation (2), 𝒑 is each point of the teaching data, 𝒑𝑖 are points which are within 𝑟1 from 𝒑, 𝑁 is number of 𝒑𝑖 . By using equation (2), the proposed method calculates the covariance matrix 𝑪 . And then, the proposed method calculates the curvature 𝐾 from eigen values𝜆0 , 𝜆1 and 𝜆2 of the covariance matrix 𝑪 by equations (1). Secondly, from 3-dimensional data shown in figure 3 (a), the proposed method extracts feature points which have a curvature higher than a threshold 𝑡ℎ to extract noticeable feature points as shown in figure 3 (b). Finally, to reduce data volume of feature points, the proposed method scans all data points with a spherical region of radius 𝑟2 and chooses one feature point which has the highest curvature in extracted feature points which are in the sphere region as shown in figure 3 (c). 2.3 List generating The proposed method generates the list of distances and angles as relationships between extracted feature points. To generate the list of relationships, firstly, the proposed method sorts extracted feature points (Figure 4 (a)) in descending order of curvature as shown in table 2. Secondly, the proposed method generates list elements by applying the rule. The rule consists of step 1 based on a highness of a curvature and step 2 based on a neighboring distance. The step 1 generates a list element in high curvatures of feature points. High curvatures are robust to noise and express characteristic

139


regions of the object. For this reason, the proposed method is able to generate unique list elements. Therefore, in recognizing the object, the proposed method is able to be less mismatching of list elements. The step 2 generates a list element in neighboring feature points from a feature point. If occlusion occurs on the target object, the proposed method is easy to match list elements of unoccluded regions with list elements generated by step 2. Therefore, the proposed method is able to recognize robust to occlusion. Here, we describe a rule as follows with figure 4 and table 2. < Rule > [Step 1] I. The proposed method selects a feature point 𝒂 which consists of 3dimensional coordinates points as a focus point and has the highest curvature in table 2. II. The proposed method selects two feature points 𝒃 and 𝒄 which consist of 3-dimensional coordinate points and which are primary and secondary next to the focus point 𝒂 in table 2. III. The proposed method obtains relationships, which are𝑙𝑎𝑏 , 𝑙𝑎𝑐 and 𝜃1 shown in figure 4, between the focus point 𝒂 and feature points 𝒃 and 𝒄 respectively. 𝑙𝑎𝑏 and 𝑙𝑎𝑐 are distances from the focus point 𝒂 to feature points 𝒃 and 𝒄 respectively. 𝜃1 is an angle between two lines which are from 𝒂 to feature points 𝒃 and 𝒄 respectively. IV. The proposed method registers above 𝒂, 𝒃, 𝐜 , 𝑙𝑎𝑏 , 𝑙𝑎𝑐 , 𝜃1 and 𝐾𝑎 , 𝐾𝑏 , 𝐾𝑐 (curvature of each feature point) as a list element shown in table 3. Table 3.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Table 2. An example of sorted feature points in descending order on curvature.

(a) An example of feature point extraction.

(b) The relationship in (c) The relationship in feature point based on feature point based on rule 1. rule 2. Figure 4. Examples of generating of list elements.

[Step 2] I.

The proposed method selects a feature point 𝒂 as a focus point which has the

Examples of list elements.

140


highest curvature in table 2. II. The proposed method selects two feature points 𝒅 and 𝒆 which are the closest from the focus point 𝒂. III. The proposed method obtains relationships between the focus point 𝒂 and feature points 𝒅 and 𝒆 which consist of 3-dimensional coordinates points respectively. 𝑙𝑎𝑑 and 𝑙𝑎𝑒 are distances from the focus point 𝒂 to feature points 𝒅 and 𝒆 respectively. 𝜃2 is an angle between two lines which are from 𝒂 to feature points 𝒅 and 𝒆 respectively. IV. The proposed method registers above 𝒂, 𝒅, 𝒆 , 𝑙𝑎𝑑 , 𝑙𝑎𝑒 , 𝜃2 and 𝐾𝑎 , 𝐾𝑑 , 𝐾𝑒 (curvature of each feature point) as a list element shown in table 3. Finally, the proposed method applies these rules to all feature points (𝒂 ~ 𝒉 which consists of 3-dimensional coordinates points) from top in table 2. 2.4 Candidate Regions Extracting To reduce processing time, the proposed method extracts the candidate regions of the target object in the scene data. In the proposed method, we assume that the target object is on a table or a floor in domestic environment. Figure 5 shows an example scene in which there are three objects on a floor. As shown in figure 5, the scene data mainly include a plane region of the table or the floor. Therefore, the proposed method firstly deletes the plane region by applying a plane detection method using

Figure 5. The assumed scene data.

Figure 6. The result of deleting a plane, and classifying each object. Rectangular region

Figure 7. The rectangle region of an object.

RANSAC [9] then the proposed method excludes (the 𝐾𝑚1plane. − 𝐾𝑠1 Secondly, < 𝑡ℎ𝐾 )⋀ the proposed method uses to cluster ( 𝐾the −clustering 𝐾 < 𝑡ℎmethod )⋀ each object as shown in figure 6. Finally, the proposed method calculates volume 𝑉c of each rectangular region of a clustered object as shown in figure 7, and calculates number of points 𝑁𝑐 of each clustered object. And then, the proposed method extracts regions of clustered objects which satisfy (3) (𝑉c ≤ 𝑉m )⋀(𝑁𝑐 ≤ 𝑁𝑚 ) as the candidate regions. Where, 𝑉m is volume of the rectangular region of the teaching data and 𝑁𝑚 is number of points of the teaching data. 2.5 List Matching In the list matching process, the proposed

Table 4. An example of the list element. (a) An example of the list element of the target object.

(b) An example of the list element of the each cluster.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

141


To recognize the target object in the scene data, the proposed method applies the rigid registration to the teaching data as shown in figure 8. Firstly, the proposed method fits the teaching data to each clustered object (Figure 9) by calculating the optimum rotation matrix 𝑹 and the translation vector 𝒕 from associated feature points which are mentioned in the list matching process (section 2.5). Secondly, the proposed method calculates a corresponding rate 𝑀 between a fitted teaching data and each clustered object by using

Figure 8. Illustration of rigid registration. Cluster 1

Cluster 2

Cluster 3

𝑁

𝑠𝑐𝑜𝑟𝑒 = ∑ 𝑓 (min{𝑑𝑖𝑠𝑡𝑖𝑗 |1 ≤ 𝑗 ≤ 𝐿}) , 𝑖=1

1 (𝑥 ≤ 𝒕𝒉𝑐 ) 𝑓(𝑥) = { 0 (𝑥 > 𝒕𝒉𝑐 ) , 𝑑𝑖𝑠𝑡𝑖𝑗 = ‖𝒑𝑖 − 𝒒𝑗 ‖,

The fitted teaching data Figure 9. Illustration of the transformed target object in each cluster.

method matches the list of the teaching data and the list of each cluster data which is extracted in the candidate regions extraction (section 2.4) to associate feature points of the teaching data and each cluster data. As shown in table 4, a list element has curvatures of selected three feature points, distances from feature point ① to feature point ② and feature point ③ respectively, the angle between each feature point. Then, the proposed method matches these lists and associates feature points of each element of these lists by finding list elements which satisfy ( 𝐾𝑚1 − 𝐾𝑠1 < 𝑡ℎ𝐾 )⋀ ( 𝐾𝑚2 − 𝐾𝑠2 < 𝑡ℎ𝐾 )⋀

( 𝑙𝑚13 − 𝑙𝑠13 < 𝑡ℎ𝑙 ), 𝜃𝑚 − 𝜃𝑠 < 𝑡ℎ𝜃 .

𝑠𝑐𝑜𝑟𝑒 𝐿

∙ 100.

(8)

Where, 𝑁 is the number of points of the teaching data. 𝐿 is the number of points of each clustered object. 𝒑𝑖 is each point of the fitted teaching data. 𝒒𝑗 is each point of the each clustered object. The proposed method counts a number of 𝒑𝑖 which are within a threshold 𝒕𝒉𝑐 which is 1 [mm] of 𝒒𝑗 by the equation (7) as a score. And then, the proposed method calculates the corresponding rate 𝑀 based on the score by equation (8). Finally, the proposed method selects a clustered object which has the highest corresponding rate.

(4)

3 EXPERIMENTS

(5)

In this section, to evaluate effectiveness of the proposed method, we compare the proposed method with conventional methods using the SHOT and the Spin Image about five properties

( 𝐾𝑚3 − 𝐾𝑠3 < 𝑡ℎ𝐾 ), ( 𝑙𝑚12 − 𝑙𝑠12 < 𝑡ℎ𝑙 )⋀

𝑀=

(7)

(6)

Where, 𝑡ℎ𝐾 , 𝑡ℎ𝑙 and 𝑡ℎ𝜃 are thresholds in equations (4), (5) and (6). (a) Spray.

2.6 Rigid Registration

ISBN: 978-1-941968-26-0 ©2015 SDIWC

(b) Pack.

(c) Cup noodle.

Figure 10. Overviews of objects and 3D data of objects in the experiment.

142


Table 5. Parameters of the proposed method.

(a) Occlusion from top side (10%).

(b) Occlusion from top side (40%).

(c) Occlusion from top side (70%). (a) Corresponding rate of the spray occluded from the top side.

(d) Occlusion (e) Occlusion from (f) Occlusion from bottom side bottom side from bottom side (40%). (10%). (70%). 70%)

(g) Occlusion (h) Occlusion (g) Occlusion from right side from right side from right side (10%). (40%). (70%). Figure 11. The examples of occlusion scene of the spray.

mentioned in section 1 as follows.  Robustness against occlusion  Fast recognition  Pose estimation with high accuracy  Coping with erroneous correspondences  Recognizing objects in different aspect ratio 3.1 Objects Recognition in Occlusion Scenes In this experiment, we compared the proposed method with conventional methods using the SHOT and the Spin Image to evaluate about three properties as follows.  Robustness against occlusion  Fast recognition  Pose estimation with high accuracy

ISBN: 978-1-941968-26-0 ©2015 SDIWC

(b) Corresponding rate of the spray occluded from the bottom side.

(c) Corresponding rate of the spray occluded from the right side. Figure 12. The result of the spray occluded from 3 directions.

We used three actual objects as recognition targets which are usually in domestic environment and obtained those 3-dimensional data with Kinect as shown in figure 10. In Figure 10, (a) shows a spray, (b) shows a pack and (c) shows a cup noodle. The spray has various curvatures because of the spray has many feature points. The pack such as a box has low and high curvatures. The cup noodle

143


(a) Corresponding rate of the pack occluded from the top side.

(a) Corresponding rate of the noodle occluded from the top side.

(b) Corresponding rate of the pack occluded from the bottom side.

(b) Corresponding rate of the noodle occluded from the bottom side.

(c) Corresponding rate of the pack occluded from the right side.

(c) Corresponding rate of the noodle occluded from the right side. Figure 14. The result of the noodle deleted from 3 directions.

Figure 13. The result of the pack occluded from 3 directions.

such as a cylinder has constant curvatures. Parameters of the proposed method are shown in table 5. To generate occlusion scenes, we delete part of each 3-dimensional object data from 3-directions (top, bottom and right side) by 10% each of point number of each 3dimensional object data as shown in figure 11. To evaluate a pose estimation accuracy of a target object, we use the corresponding rate 𝑀 between the target object fitted by using the optimum rotate matrix 𝑹 and the translation vector 𝒕 mentioned in the rigid registration process (section 2.6). To calculate the corresponding rate 𝑀 as a pose estimation

ISBN: 978-1-941968-26-0 ©2015 SDIWC

accuracy, we use the equation (7) and (8) with 𝒕𝒉𝑐 which is 10 [mm]. In case that, the corresponding rate is high, that means methods estimate the pose of a target object with high accuracy. On the contrary, in case that, the corresponding rate is zero, that means methods mismatch the target object. The reported processing time is obtained using Intel(R) Table 6. The result of average processing time.

144



Core(TM) i5 3.1GHz with 8.0 GB of main memory. Figure 12 shows the result of occlusion scenes for the spray object. Figure 13 shows the result of occlusion scenes for the pack object. Figure 14 shows the result of occlusion scenes for the cup noodle object. As shown figure 12, figure 13 and figure 14, the proposed method was able to recognize objects nearly equal to conventional methods in occlusion scene, the corresponding rate of the proposed method is nearly equal to the corresponding rate of conventional methods. In addition, as shown in table 6, a processing time of the proposed method was nearly equal to the SHOT and was faster than a processing time of the Spin Image.

(a) The result of the SHOT.

(b) The result of the Spin Image.

(c) The result of the proposed method. Figure 15. The result of each method.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

From these results, we consider that the proposed method has the robustness against occlusions because the proposed method is able to match feature of the target object and feature of unoccluded scene data by using the curvature which is calculated as a local feature. In addition, we consider that the processing time of the proposed method is fast because the proposed method uses sparse feature point. Furthermore, we consider that the proposed method is able to estimate a pose of a target object with high accuracy because the proposed method uses is not only the feature points but also relationships between feature points. 3.2 The Experiment in Recognition of Objects which have the Same Feature but which are not the Same Object To qualitatively evaluate a coping with erroneous correspondences in the proposed method, we compared the proposed method with conventional methods using the SHOT and the Spin Image. As target objects which have the same feature in a local region but which are not the same object, we prepared the spray and cup noodle mentioned in section 3.1. Parameters of the proposed method are shown in table 7. In this experiment, we generate the teaching data from the cup noodle and apply the proposed method, SHOT, Spin-Image to a scene data where there are the spray and the pack mentioned in section 3.1. Figure 15 shows results of these methods. In figure 15, lines show corresponded feature points of the teaching data and the scene data. As shown in figure 15 (a) and (b), the SHOT and the Spin Image have a lot of corresponded feature points between the cup noodle with the spray. That means these two methods were not able to distinguish an object which has the same feature but which are not the same object and

145



(a) 350-ml can. (b) 500-ml can. Figure 16. Overviews of objects and point clouds used in the experiment.

method has no corresponded feature points. That means the proposed method was able to distinguish the cup noodle and the spray because the proposed method uses not only feature points but also relationships between feature points. From these results, we consider that the proposed method is an effective recognition method for coping with erroneous correspondences. 3.3 The Experiment in Recognition with Change in Aspect Ratio

(a) The result of the SHOT.

(b) The result of the Spin Image.

(c) The result of the proposed method. Figure 17. The result of each method.

misrecognized the spray as the cup noodle. By contrast, as shown in figure 15 (c), the proposed

ISBN: 978-1-941968-26-0 ©2015 SDIWC

To qualitatively evaluate an ability of the proposed method in recognitions of objects which have different aspect ratio, we compared the proposed method with conventional methods using the SHOT and the Spin Image. As target objects which have different aspect ratio, we prepared a 350-ml can and a 500-ml can as shown in figure 16. Parameters of the proposed method are shown in table 8. In this experiment, we generate the teaching data from the 350-ml can and apply the proposed method, the SHOT and the Spin-Image to a scene data in which there is the 500-ml can. Figure 17 shows results of these methods. In figure 17, lines show corresponded feature points of the teaching data and the scene data. As shown in figure 17 (a) and (b), the SHOT and the Spin Image have a lot of correspondence feature points. That means these two methods were not able to distinguish the difference of aspect ratio and misrecognized the 500-ml can as the 350ml can. By contrast, the proposed method has no corresponded feature points. That means the proposed method was able to distinguish the different aspect ratio by relationships between feature points and correctly recognized the 500ml can as not the 350-ml can. From these results, we consider that the proposed method is

146


an effective recognition method for different aspect ratio objects. 4

REFERENCES [1]

S. Sugano, T. Sugaiwa and H. Iwata, “Vision System for Life Support Human-Symbiotic-Robot”, The Robotics Society of Japan, 27 (6), pp. 596-599, 2009.

[2]

T. Odashima, M. Onishi, K. Tahara, T. Mukai, S. Hirano, Z. W. Luo and S. Hosoe, “Development and evaluation of a human-interactive robot platform “RI-MAN”” The Robotics Society of Japan, 25 (4), pp. 554-565,2007.

[3]

Y. Jia., H. Wang.,P. Sturmer and N. Xi, “Human/robot interaction for human support system by using a mobile manipulator”, ROBIO, pp. 190195, 2010.

[4]

F. Tombari and S. Salti, “Unique signatures of histograms for local surface description”, ECCV, pp. 356-369, 2010.

[5]

F. Tombari, S. Salti and L. D. Stefano, “A Combined Texture-Shaped Descriptor for Enhanced 3D Feature Matching”, ICIP, pp.809-812, 2011.

[6]

A. E. Johnson, “Spim-Images: A Repersemtation for 3-D Surface Matching, doctoral dissertation”, The Robotics Institute, Carnegie Mellon Univ., 1997.

[7]

A. E. Johnson and M. Hebert ”Using Spin Images for Efficient Object Recognition in Cluttered 3D Scenes”, Trans. IEEE Pattern Analysis and Machine Intelligence, 21,5, 1999.

[8]

C.Conde, R.Cipolla, L.Rodriguez-Aragon, A.Serrano and E.Cabello, “3D facial Feature Location with Spin Images”, IAPR Conference on Machine Vision Apprications, pp.418-421, 2005.

[9]

M. A. Fischler and R. C. Bolles, “Random Sample Consensus: A Paradigm for Model Fitting with Applications to Image Analysis and Automated Cartography,” Comm. of the ACM, 24 (6), pp.381395, 1981.

CONCLUSION

In this paper, we proposed the 3-dimensional object recognition method which has five properties as follows using relationships of distances and angles in feature points for the human support robot.  Robustness against occlusion  Fast recognition  Pose estimation with high accuracy  Coping with erroneous correspondences  Recognizing objects in different aspect ratio From experimental results for recognizing objects which have occlusion regions, we considered that the proposed method has the robustness against occlusions and the processing time of the proposed method is faster. In addition, we considered that the proposed method is able to estimate a pose of objects with high accuracy. Furthermore, in experimental results for recognizing objects which have the same features but which are not the same object, the proposed method did not mismatch these objects and we considered that the proposed method is more effective object recognition method than conventional methods. Furthermore, in experimental results for recognizing objects which have different aspect ratio, the proposed method did not misrecognize these objects and we considered that the proposed method is more effective object recognition method than conventional methods. However, since the proposed method only uses curvatures which are calculated from a shape data of objects, the proposed method is not able to recognize objects which have the same shape and different texture. Therefore, we improve the proposed method by using not only curvatures but also color features in a future work.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

147


Disparity Map Estimation using Gabor Wavelet under Radiometric Changes Malathi. T1, M. K. Bhuyan1, C. Asokan2, and Y. Iwahori3 1

Department of Electronics and Electrical Engineering Indian Institute of Technology Guwahati, Assam-781039, India E-mail:{malathi,mkb}@iitg.ernet.in 2 Department of Electronics and Communication Engineering KSR Institute for Engineering and Technology, Tamil Nadu-637215, India E-mail:[email protected] 3 Department of Computer Science Chubu University 1200 Matsumoto-cho, Kasugai 487-8501, Japan E-mail: [email protected]

ABSTRACT 1 INTRODUCTION Stereo correspondence refers to the problem of finding the matching pixels of a 3D world point in the stereo image pairs captured at different viewpoint using two cameras. Subsequently, the obtained disparity map finds application in object detection, augmented reality and robotics. In recent years, a large number of algorithms have been developed to compute accurate disparity map. Many existing disparity map estimation methods are based on the brightness constancy assumption. Unfortunately, any lighting variations within the scene violate this assumption. In this paper, a disparity map estimation method using Gabor wavelet is proposed. The performance of the proposed method is appraised under varying synthetic illumination change and real radiometric variations. Experimental results under non-ideal radiometric conditions illustrate that the proposed method respond to the changes in the illumination and radiometric variations.

KEYWORDS Feature extraction, Gabor filter, Kuwahara filter, Cost aggregation, disparity map.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Stereo correspondence is active for more than a decade finding application in image rendering, robotics and video surveillance. Disparity map represents the horizontal displacement of all the pixels from the left image to the right image. The four major steps in stereo correspondence methods can be categorized as follows: (i) Matching cost computation, (ii) Cost aggregation, (iii) Disparity computation and (iv) Disparity refinement [1]. The matching cost computation is based on the constant intensity assumption at the matching pixels. This assumption is violated when the stereo images undergoes radiometric changes. Radiometric differences can be caused by camera settings, vignetting and image noise. Another source of radiometric differences is due to non-Lambertian surface for which the amount of light reflected from the surface depends on the viewing angle. Acquiring images of static scene at different times may change the strength or positions of the light sources which forms another source of radiometric changes. Due to reasons listed

148


above, real-world stereo application requires the stereo matching is robust to radiometric changes. Without additional computation, more robust cost functions can compensate for nonideal radiometric conditions. Hence, in this paper we analyze the robustness of the proposed method to radiometric changes. 2 RELATED WORKS The basic matching cost function used in stereo correspondence are sum of absolute difference (SAD), sum of squared difference (SSD) and normalized cross correlation (NCC) [2]. In addition to this, nonparametric transforms such as rank and census transform [3], mutual information [4] and permeability [5] are also used for matching cost computation. Stereo matching algorithms are based on the wellknown assumption that corresponding pixels have similar intensity values. In real scenario, various factors affect this assumption and prevent the corresponding pixels from having similar intensity values. One such factor is the radiometric change which includes different illumination and camera exposure change. As opposed to SAD and SSD, normalization in NCC takes care of gain difference i.e., a multiplicative change between the stereo images [6]. Zero- mean SAD (ZSAD), ZSSD and ZNCC accounts for a constant offset (an additive change). Another possibility of reducing bias offset is by filtering the stereo images before matching using mean filter, computing magnitude of first or second derivative. Miserably, use of these filters result in blurred disparity map. Rank and census transform produce steady response under all radiometric changes, this is due to the fact that

ISBN: 978-1-941968-26-0 ©2015 SDIWC

nonparametric transformation depends on the relative ordering of the intensity values. Mutual information (MI) facilitates correspondence between input pairs by computing the joint histogram of pixel values between the stereo images. MI-based costs are robust to globally transformed images but fail to handle local radiometric changes. In [7], the concept of adaptive normalized cross-correlation (ANCC) was proposed. In this method, color insensitive information is extracted and then the new measure ANCC is used for stereo matching. This method is robust to local and global radiometric variation but suffers from multiple illumination conditions. To produce accurate disparity map under varying radiometric conditions, Heo et al proposed data cost which is a combination of mutual information and SIFT descriptor and segment-based plane plane-fitting [8]. Mean-shift algorithm increases the computational burden of the algorithm. Evaluation of these matching costs for simulated and real radiometric differences can be found in [6]. The main contributions in this paper are as follows: (1) Local features for matching cost computation are extracted using Gabor wavelet. (2) Cascaded Kuwahara and median filters are used for cost aggregation. (3) The robustness of the proposed method to synthetic and real radiometric changes is evaluated. This paper is organized as follows: Section 3 describes in detail the proposed stereo correspondence method, section 4 shows the experimental results and finally, conclusion in section 5.

149


Figurre 1: Block diaagram of the prroposed disparrity map compuutation methodd.

3 PROPOSED METHO OD

3.1 M Matching C Cost Compu utation

M Most of thee existing sttereo matchiing methodss hhave the folllowing four steps namely - (i) M Matching cost com mputation, (ii) Costt c n aggregation,, (iii) Disparity map computation and (iv) Disparity refineement. Block k diagram off tthe proposed disparity map estimattion method d along with the t intermed diate results is shown in n F Figure 1. A brief explan nation of thesse steps in tthe proposed d method is as follows: Local L Gaborr w wavelet features fe reeduced by y Principall Component analysis (PCA) iss used to o accomplish matching g cost computation, c , K Kuwahara and median n filter perrforms costt aggregation,, disparity map is co omputed by y w winner-take-all (WTA) techniques and finally,, oocclusion detection d an nd filling followed f by y ddisparity refinement. All these steps aree ddescribed in n detail in th he remaining g part of thiss section.

Featture of a pixxel in the left ft image is coompared withh the featuree of a pixel iin the right iimage to findd the matchinng cost for m maximum allowable dispparity values. Feature maay be intensiity, color or ttexture. In this paper, Gabor waavelet is empployed to extract locaal features.. Gabor wavvelet is a w widely used for texturee feature many vision extraaction in computer appllications. The motivation behind usinng Gabor waavelet for featuure extractioon is as folloows [9]:

ISBN: 978-1-941968-26-0 ©2015 SDIWC

 Simple cells in thhe visual coortex of mammallian brains ccan be best m modeled by Gaboor function.  Image perception by humann visual system iis similar too image anaalysis by Gabor fuunction.

150


(a)

(b)

(c))

(d))

F Figure 2: Gaabor extracted features. (a) Input teddy image, i (b) im mage representeed using real coefficients, (c) image rrepresented using imaginary coefficients an nd (d) image reepresented usinng magnitude innformation.

Gabor fun nctions are Gaussian modulated d ccomplex sin nusoids given n by [11]

1 184x2y2  ix k22  g(x, y)  e e e   2

(1)

Gabor waveelet is referrred as a cllass of self-similar funcctions generrated by thee process off oorientation and the scaaling of thee 2D Gaborr function giv ven by

gmn (x, y)  amg(xa , ya ), a 1 xa  am (x cos  y sin ) and

(2)

xa  am (xsin  y cos ) w where,   n , m and n are two inteegers and k k

iis the total number n of orrientations. Consider an n image I of size P  Q . In order to o find the feaature vector for the pix xel I (i , j ) , a ccertain neig ghborhood N (i , j ) of siize u  v iss

F (i, j)  cooncat( mn (i, j))

 mn (i, j)  N (i, j)  reaal (gmn )

(3)

wheere,  is tthe convoluution operattion and peration. con cat denotes the concaatenation op Thiss procedure is repeated for all the ppixels in the image. The dimensionaality of the obtained featuures is redu duced by PC CA [12]. F Figure 2 show ws the extraccted feature for the teddyy image. Mattching cost is computedd by compaaring the pixeels in the left image wiith the pixells in the righht image aloong the horrizontal linee for all posssible disparity values. T The more sim milar the pixeels are the leesser is the ccost value. M Matching costt is a 3D volume with cost valuess for all pixeels at differennt disparity vvalues. 3.2 C Cost Aggreggation

cconsidered. Here, (i , j ) is the pixel coordinates.. T This patch is convolved with the Gabor filterr kkernel g mn for fo different orientations o and scaling.. Gabor waveelet is a com mplex filter and a here wee hhave used only o the real part of Gab bor filter forr feature exttraction. Th he featuress are then n extracted by b concateenating thee obtained d ccoefficients given by: Figu ure 3: Kuwaharra filter subreggions.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

151


Cost aggregation is the process of smoothing or averaging of computed matching cost for a particular disparity value. In the proposed method we have used cascaded Kuwahara [13] and median filter for cost aggregation. Edge preserving property and run time of O (1) makes Kuwahara filter suitable for cost aggregation. Median filter is used to remove the blocking artifacts produced by Kuwahara filter. Kuwahara filter performs smoothing by dividing the neighborhood of length 2a for each pixel into four subregions namely Region 1 (Q1), Region 2 (Q2), Region 3 (Q3) and Region 4 (Q4) as shown in Figure 3 which is given by Q1 (i, j )  i, i  a    j , j  a   Q2 (i, j )  i  a, i    j , j  a   Q3 (i, j )  i  a, i    j  a, j  Q (i, j )  i, i  a  j  a, j      4

(4)

where the symbol " " denotes the Cartesian product. Local mean mz (i, j ) and variance  z (i, j ) are computed for each subregion Qz , z  1, , 4 . The mean of the subregion which has minimum variance among the four regions is assigned to the center pixel (i, j ) , formulated as

 (i, j )   mz (i, j ) f z (i, j )

(5)

z

where 1,  z (i, j )  min k  k (i, j ) f z (i, j )   0, otherwise

taking the index of the minimum value in the aggregated cost of the corresponding pixel. Mathematically, the disparity d u of a pixel u is given by

du  arg min CA(u, d )

where, CA(u , d ) is the aggregated matching cost of a pixel u at the disparity d . Here, D denotes the set of allowed disparity values. 3.4 Disparity Refinement Occluded pixels are filtered out by the left-right consistency check i.e., another disparity map is extracted by keeping the right image as the reference image. Subsequently, pixels in the left disparity map are compared with the corresponding matching points in the right disparity map. Apparently, it is done to check whether both the disparity maps carry the same disparity value. If the test fails, then the particular pixel is marked as occluded. In occlusion filling step, the disparity d u of the occluded pixel u is assigned a value of min( d l , d r ) , where dl and d r are the first valid left and right neighbors of the pixel u . Disparity refinement is performed by a constant time weighted median filter [14]. The weights are calculated by the guided filter. The weights W (i, j ) are given by W (i, j ) 

1

2

where I i ,



1   I       U  T

 :( i , j )

i

I j and

covariance matrix 3.3 Disparity Computation The disparity map is obtained by determining the disparity d u of all the pixels u  (i , j ) in the reference image. This is accomplished by

ISBN: 978-1-941968-26-0 ©2015 SDIWC

(6)

d D





 are

 and

1

I

j

  



(7)

3 1 vectors.

The

the identity matrix

U have a size of 3  3 . Again,  denotes the

number of pixels in the window  and  is a smoothness parameter.

152


4 EX XPERIMEN NTAL RESULTS

Allgorithm for f dispariity map compu utation

The proposedd method is evaluaated on Midddlebury steereo datasetss [1], [15]. All the expeerimental reesults shownn in this paaper are evalluated for thhe error thresshold of 1. IIn all the resuult shown, N Nocc, all andd disc repreesent the perccentage of bad pixeels (pixels having dispparity valuess that deviattes from thee ground truthh by +/- 1) in the noon-occludedd region, entirre image and disccontinuous regions resppectively. Inn this sectioon, we exam mine the perfformance oof the propposed methhod for syntthetic illuminnation changges. To expllore this, the left image is unchangged while tthe right mination imagge undergooes the synnthetic illum variaations. Thiss experimennt is carriedd out for bothh simulated illuminationn changes and real radioometric variiations.

1. Matching cost computation c n 1.1 Imag ge patch is convolved with w the reall Gab bor filter witth different orientationss and scaling F (i, j)  concat c ( mn (i, j))  mn (i, j)  N (i, j)  real( gmn ) ng extracteed featuress, computee 1.2 Usin matcching cost for f all possib ble disparity y valu ues. 2. Costt aggregatio on 2.1 Smo ooth matchin ng cost for all disparity y valu ues  (i , j )   m z (i , j ) f z (i , j ) z

1,  z (i, j )  min n k  k (i, j ) f z (i , j )   0, otherwise 1,  z (i, j )  min k  k (i, j ) f z (i, j )   o 0, otherwise

3. Disp parity computation 3.1 Afteer cost aggregation, disparity valuee corrresponding to minim mum costt consstitutes the disparity d map p.

(a)

(b)

(c)

d u  arrg min CA(u , d ) d D

4. Occlluded pixelss detection and a filling 4.1 Occluded pixelss detected by b left-rightt conssistency checck. 4.2 Occlusion fillin ng is done by b assigning g miniimum of left and rightt neighbor’ss disp parity values.. 5. Disp parity refineement 5.1 Disp parity map refinement r by b weighted d med dian filter. W (i , j ) 

1

2



 :( i , j )

1   I       U  T

i





1

I

A Algorithm 1: Proposed algorrithm for dispaarity map ccomputation.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

j

  

(d)) (e) Figu ure 4: Syntheetic illuminatiion changes: (a) Input Gain (0.5); imagge; (b-e) syntheetically variedd images; (b) G (c) B Bias (30); (d) Gamma (0.5); (e) Vignettting effect (0.5) .



4.1 Simulated iillumination n Changes: In realtimee scenario, illuminationn changes may be eitheer global or local. Globaal change in turn can be eeither linearr or nonlineear. Gain annd offset channge constituutes the linnear global change wheereas nonlinnear changee is represeented by gam mma changge. Despitee of thiss local radioometric chaange is dennoted by viignetting effecct.

153


Locaal radiometrric change  Vignettinng effect We use a singlee formula foor global raddiometric channge that includes both llinear and nnonlinear channges [16]. M Mathematicallly, it is given by: f   m f Ii  a f   I o  uint 8  255    , m f and  f  0   255    (8)

wheere, I i and Io are inputt and synthhetically (a)

(b)

varieed output im mages respeectively. m f , a f and  f arre the multiiplicative, aadditive and d gamma

factoors respecttively. Figuure 4 shoows the diffeerent syntheetic illuminaation changee applied to aan input im mage and F Figure 5 shows the quanntitative com mparison off its effectss in the obtaained disparrity map. F For global ssynthetic channge such aas multipliccative and gamma factoor, the errorr increases with the am mount of channge in intenssity values w whereas for additive factoor the errorr remains allmost constaant. This increease in errror is duee to the inherent propperties of the convolutioon operationn. Figure 6 shhows the efffect of vignnetting on ddisparity mapp computattion. Similar to the global radioometric diffference, in loocal differennces also erro r increases with channge in raddiometric variaations.

(c) ( F Figure 5: Gllobal syntheticc illumination n changes: (a)) N Nonlinear-gam mma; (b) liinear-additive; (c) linear-m multiplicative.

Global simu ulated changee  Linear Gain chaange (multipllicative) Offset orr bias changee (additive)  Nonllinear Gamma change c

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Figu ure 6: Radiomeetric differencees - Vignetting effect.

154


the light sourcees [17, 18]. So, totallyy we havee nine differrent images which reflecct the scennarios mentiioned abovee. Figure 7 aand 8 show ws the dissparity maap for diffferent expoosures annd illuminnation chhange resppectively aloong with thee ground truuth. It show ws the perrformance oof the propposed methhod to real rradiometric variations. H Here, Exp 0/1 denotes left image is from exposure 0 w whereas rightt image is ffrom exposuure 1. Sim milarly, Illu3//2 denotes leeft image is from illum mination 3 w whereas righht image is from illum mination 2. T Table 1 and 2 gives the m mean squaare error for the results sshown in Figures 7 annd 8 respectivvely.

Figure 7: 7 Real radiiometric diffeerences using g different ex xposures.

quare erro or for reall Table 1 Mean sq radiomeetric change - Exposuree Left/rig ght image 0 1 2

0

1

2

662 1080 0 2945 5

1014 4 677 3591

3467 2864 1187

ons: Though h 4.2 Reall Radiometrric variatio Middlebu ury datasets used u aree radiomettrically clean n and does not requiree robust stereo match hing costs (except thee d syntheticc changes) but the real-world applicatio ons may su uffer from radiometricc changes. Hence, we analyze thee robustnesss m to radiometricc of the proposed method ns. Stereo datasets d con nsidered forr variation this anallyze are acq quired at thrree differentt exposurees and underr three confiigurations off

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Figu ure 8: Real radiometric differences using differrent illumination.

Tab ble 2 Mean square error for real radiiometric change - Ligh hting Leeft/right iimage 1 2 3

1

2

3

677 1267 725

795 563 732

6699 11189 6661

155


images with different illuminations and cameras,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 35, no. 5, pp. 1094–1106, 2013.

5 CONCLUSIONS In this paper, we proposed a disparity map estimation method. In addition to this, we have also evaluated the performance of the proposed method for synthetic (both local and global) and real radiometric differences. These experimental result shows that the proposed method produces better results under radiometric difference.

[9]

T. S. Lee, “Image representation using 2D Gabor wavelets,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 18, pp. 959-971, 1996.

[10]

J. G. Daugman, “Uncertainty relation for resolution in space, spatial frequency and orientation optimized by two-dimensional visual cortical filters,” Journal of the Optical Society of America A, vol. 2, pp. 1160-1169, 1985.

[11]

S. Bhagavathy, J. Tesic and B. S. Manjunath, “On the Rayleigh nature of Gabor filter outputs,” IEEE Int. Conf. on Image Processing, 2003, pp. 745-748.

[12]

M. K. Bhuyan and Malathi. T, “Review of the Application of Matrix Information Theory in Video Surveillance,”in Matrix Information Geometry, Frank Nielsen and Rajendra Bhatia, Ed. Springer, 2012, pp.293-321.

[13]

G. Papari, N. Petkov and P. Campisi, “Artistic Edge and Corner Enhancing Smoothing,” IEEE Trans. Image Processing, vol. 16, pp. 2449-2461, 2007.

[14]

Z. Ma, K. He, Y. Wei, J. Sun and E. Wu, “Constant Time Weighted Median Filtering for Stereo Matching and Beyond,” Int’l Conf. Computer Vision and Pattern Recognition, 2013, pp. 1-8.

[15]

D. Scharstein and R. Szeliski, “High-accuracy stereo depth maps using structured light,” Int’l Conf. Computer Vision and Pattern Recognition, 2003, pp. 195-202.

[16]

M. A. Mohamed, H. A. Rashwan, B. Mertsching, M. A. Garcia and D. Puig, “Illumination-robust optical flow approach using local directional pattern,” IEEE Trans. Circuits Syst. Video Technol., vol. 24, no. 9, pp.1499-1508, 2014.

[17]

H. Hirschm̈ uller and D. Scharstein, “Evaluation of Cost Functions for Stereo Matching,” Int’l Conf. Computer Vision and Pattern Recognition, 2007, pp. 1–8.

[18]

D. Scharstein and C. Pal, “Learning conditional random fields for stereo,” Int’l Conf. Computer Vision and Pattern Recognition, 2007, pp. 1–8.

REFERENCES [1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

D. Scharstein and R. Szeliski, “A taxonomy and evaluation of dense two-frame stereo correspondence algorithms,” Int’l J. Computer Vision, vol. 47, pp. 7-42, 2002. M. Z. Brown, D. Burschka and G. D. Hager, “Advances in Computational Stereo,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 25, pp. 993-1008, 2003. R. Zabih and J. Woodfill, “Non-parametric Local Transforms for Computing Visual Correspondence,” European Conf. on Computer Vision, 1994, pp. 151-158. H. Hirschmuller, M. Buder and I. Ernst, “Stereo processing by semiglobal matching and mutual information,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 30, pp. 328-341, 2008. C. C. Pham, V. D. Nguyen and J. W. Jeon, “Efficient spatio-temporal local stereo matching using information permeability filtering,” IEEE Int. Conf. on Image Processing, 2012, pp. 29652968. H. Hirschm̈ uller and D. Scharstein, “Evaluation of stereo matching costs on images with radiometric differences,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 31, no.9, pp. 1582-1599, 2009. Y. S. Heo, K. M. Lee, and S. U. Lee, “Robust stereo matching using adaptive normalized crosscorrelation,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 33, no. 4, pp. 807–822, Apr. 2011. Y.S. Heo, K.M. Lee, S.U. Lee, “Joint depth map and color consistency estimation for stereo

ISBN: 978-1-941968-26-0 ©2015 SDIWC

156


The Effect of Information Systems in the Information Security in Medical Organization of Shaharekord

Leila Rahmani Samani Msc Student of Government Management Isfahan (Khorasgan) Branch, Islamic Azad University, Isfahan, Iran [email protected] Mohmmadreza Soltanaghaei Department Computer Assistant Professor

Isfahan (Khorasgan) Branch, Islamic Azad University, Isfahan, Iran [email protected]

Abstract

1. Introduction

Today`s spread of sciences and emergence of modern technologies necessitates information exchange. As a result security of information systems should be considered more seriously. The purpose of the present study is to determine the effect of information systems in organizational information system. Accessibility, privacy, authentication, undeniability, and comprehensiveness are the five effective dimensions of information security that have been analyzed using a researcher made questionnaire and were analyzed in a statistical population of 165 personnel of medical organization of shahrekord and sample of 115 people using a simple randomization method. The validity of the questionnaire was analyzed by face validity and the reliability was analyzed by internal consistency (=0.92). The data were analyzed using SPSS 19 and the results showed that information systems influence all dimensions of information security i.e. accessibility, privacy, comprehensiveness, undeniability and authentication more than average and in general the role of information systems in information security of medical organization of Shahrekord is more than average.

Information security means protecting information and information systems away from unauthorized activities. These activities include accessibility, use, disclose, copying and recording, destroying, altering and manipulating [1]. The goals of information systems are as follows:

Keywords:

Comprehensiveness: Ensuring that the information is accurate and complete.

accessibility, privacy, comprehensiveness, undeniability, authentication.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

a. Collecting and storing data from various sources in an integrated and coherent format b. The ability to rapidly respond to the needs of applicants c. the possibility of exchanging information between different centers d. Ability to prepare comparative reports e. Providing necessary information tools for planning and quick and easy decisionmaking [2]. To have a secure environment some main factors should be present. These factors are:

Privacy: Ensuring that information is usable only by persons or organizations

157


authorized and there will be no disclosure of information for unauthenticated people. Identification and authentication: The receiver and transmitter, both need to be sure of the identity of the other party. Accessibility: ensuring that the system is responsible for the delivery, storage and processing of data and is always available when needed by relevant people. Undeniabiliy: Neither of the parties deny its participation in the association. The common mistakes that can be made and cause a security problem with regard to some sensitive data in an organization can be divided to three groups as follows. A. common mistakes of system managers: the absence of a personal security policy, connecting systems without a proper Using information systems

configuration to the internet, trusting too much to tools, neglecting logs executing extra or unnecessary services or scripts. B. common mistakes of organization managers: the use of untrained unskilled experts, a lack of awareness about the impact of a security weakness on organizational performance, lack of funding for addressing information security, complete reliance on tools and commercial products, low investment in information security. C. common mistakes of ordinary users: violation of the organization's security policies, sending sensitive data on home computers, taking notes of sensitive data and their insecure storage, receiving files from untrusted sites and non-compliance with physical security[3].

Security of organization data

Accessibility

Privacy Authentication undeniability ComprehenAccessi bility

Figure 1. Conceptual model of study

C

Hypotheses: Main hypothesis: The use of information systems has a role in the security of organizational information of Medical Sciences organization of Shahrekord. 1. the use of information systems has a role in the undeniability of organizational

ISBN: 978-1-941968-26-0 ©2015 SDIWC

undeniability

information of Medical Authentication organization of Shahrekord.

Sciences

2. the use ofPrivacy information systems has a role in the accessibilty of organizational siveness information of Medical Sciences organization of Shahrekord. 3. the use of information systems has a role in the privacy of organizational information

158


of Medical Sciences organization of Shahrekord. 4. the use of information systems has a role in the authentication of organizational information of Medical Sciences organization of Shahrekord. 5. the use of information systems has a role in the comprehensibility of organizational information of Medical Sciences organization of Shahrekord.

2. Literature Review Organizations need to consider information security to stay safe from these harms. To improve information security in information systems in all aspects issuing a national guideline seems essential. However, the security of different organizations taking their different aspects can be considered separately in each section. For example, to ensure security in technical dimensions solutions such as the use of new software to record important changes in the system, limiting the levels of access based on user roles and duties, changing passwords periodically and periodic penetration testing to ensure software security is recommended. Using standard server room equipped with a fire alarm system, equipping devices with emergency power to prevent physical injuries and further support for managers in physical security is of the recommendations that could increase the security of devices and information. Many studies have been done in this field including: Mozani(1394) in his study entitled management information systems declared that information systems are the use of computer technology. Besides the difficulty and cost of implementing this system it can help managers spend little time to obtain more accurate and better data.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Molavi and Khenifer (in their study) of organization information systems that was conducted using a questionnaire in all manufacturing and service organizations stated that the quality of information of information systems is a consequence of organizational citizenship behavior and that altruism in relation to generosity and sense of duty has a greater impact on the quality of information system. Khaksar Haqqani and colleagues (1392) in an article entitled the analysis of security in information systems state that: The function of information systems is processing data, but sometimes it is thought that in information systems of organizations what happens is the mere simple processing of raw data from financial and non-financial events affecting the activities of the organization. Managing information systems, especially in cases where these devices are not secure is necessary, because the dangers that may threaten these devices are too much. In Companies and organizations in which the security of information systems is weak, the risk of penetration and manipulation of information in the system is high and the damage can irrecoverable. The need for the information security requires that the previsions be made by the management to ensure the security of information and make the information reliable. Mehrayin and colleagues (1392) in their analysis of the information security in information systems in hospitals among the managers of information technology in hospitals affiliated to Tehran University of Medical sciences and Shahid Beheshti and representatives of software companies totaled of 99 said that security level in hospital information systems was assessed as acceptable. However, planning for the development and implementation of the latest security policies and guidelines in all three dimensions of managerial, technical and physical in accordance with user needs and technological advances seems necessary.

159


Seifikar and collegues (1392) in their analysis of the effectiveness of information systems security model in sample population of research, commercial, manufacturing, consulting and service organizations reached to the conclusion that information is vital in organizations, advanced institutions and scientific communities. Access to the information and its suitable and quick supply have always been considered by the organizations that information has a central and crucial role for them. Organizations and institutions should provide a convenient infrastructure for their organizations and should move toward their organization and security of information in their organizations. Elahi and colleagues (1388) provided a framework for related human factors in the security of information systems in the sample population of organizations subset of ministry of communications and information technology. In this framework human factors are more effective on the security of information systems in relation to technical and tactical factors so by paying more attention to them can have a greater impact on improving the effectiveness of security and protection of information assets in organizations. Mahmoodzadeh and Rajabi (1385) in their analysis of the information systems security management in the sample of 55 of investigated the security management of information systems in 55 of computer users in three selected organizations concluded that given that the main concern of the staff is to meet the demands and tasks that are ignored, so the first thing that can be stated as problem is the lack of clearly articulated rules that are communicated to the employees. The most important thing is that the lack of written rules first inhibits employees from knowing their duties in relation to protecting organization information and second in the case of committing violations there is no authority to address it. Due to the great weaknesses

ISBN: 978-1-941968-26-0 ©2015 SDIWC

of the organization in maintaining their information assets, organizations should identify their critical assets by the implementation of information security management system and then determine their risks. Then by controlling the risks choose methods to minimize the risks. Kim et al (2015) investigated the relationship between the need for stretching / compression technology and information security management and the role of moderating regulatory pressure and came to the conclusion that the purpose of management information systems is to minimize the damage to the organizations by preventing and controlling security problems that stem from the manipulations and unexpected events. However, security is not only a technical issue but also a very complex issue including intentional or inadvertent threats both inside and outside of the organization. Durk and Choi (2014) offered the dynamic system model for information security management and concluded that secure information assets have vital importance for the organization. Although it is unlikely all the information assets of an organization to be safe and secure and providing security is expensive but it is essential for an organization to enforce security. The model includes many practical aspects of security, including attacks, detection, recovery, risk assessment and vulnerability reduction. Chang and Wang (2011) in a study entitles as sources of information systems and information security say: the importance of information security awareness has been shown by the rapid increase in investment in information security. One of the potential threats to an organization is the breach in information security. In this way system resources play a significant role. Information security level of an organization refers to people and security infrastructure and relationships so it should be considered seriously.

160


3. Methodology Methodology: the present study is an applied one because applied research are those that using the cognitive and informational context already provided by previous research are used to meet human needs. This article is descriptive in nature, by definition, descriptive studies or research are those that aim to describe the phenomena under study. Conducting descriptive research can merely aid decision-making or identifying the conditions under study. In this study the data collection method is field study survey. Field studies are scientific studies that investigate the relationship between the variables systematically and analyze the hypotheses in the real life situation of local communities, schools, factories and institutions. So it can be firmly said that the data collection method in this study is a survey and is conducted through a questionnaire. Sample population: A group of people with one or more attributes and the trait or traits are considered by the researcher. Sample may include all people, a particular type or a limited number that group. The study sample included all managers and professionals and staff who work with the information system of Medical Sciences Organization of Shahrekord totaled 165 persons. Sample size: The sample consists of a series of signs that is selected from a part or a group of a larger society in a way that the characteristics of this collection represents the quality and characteristics of the larger society. The method used in this study to determine the sample is the use of statistical methods and techniques, however, due to the fact that the investigator needs to know the information and parameters about the society he is going to select the sample from. According to the results of the questionnaires using the following formula we can calculate the sample size.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

z 2 pq 2

d n 1  t 2 pq  1   2  1 N d 

(1 )

According to this formula the sample size was assessed to be 115 persons. 140 questionnaires were distributed but 117 were returned. The data collection tools in the study are as follows: Library research and the questionnaire: to come up with the questionnaire at first literature review was used and then according to the aims of the study they were used and then consulting the advisor and counsellor the final version of the questionnaire was prepared. The reliability of the questionnaire was estimated by the advisor. There are several ways to measure the validity of the questionnaire such as testretest, parallel method, classification method and Cronbach's alpha coefficient. To assess the validity of the questionnaire in this study, Cronbach's alpha coefficient was used. This method is used to calculate the internal consistency of measurement tools such as questionnaires or tests that are used to measure different characteristics. In this tool the response to each question can take different values. The reliability value can range from -1 (no relation) to +1 (full relation). To find Cronbach's alpha scores first variance for each subset of items in the questionnaires (sub-test) should be calculated and then then total variance. Then using the following formula Cronbach's alpha can be calculated. 2  j  S j  1  r    S 2   j  1  

(2)

r : Cronbach's alpha coefficients J: the number of items in the questionnaire Sj: sub-test variance

161


S2: total test variance

significant differences between the sample and the average population. The difference between the sample and population average is 0.301. In other words, the sample average of 3.3 is greater than population average (3) by about 0.301unit. So the role of information systems in the information security of Medical Sciences organization is assessed to be more than average.

4- Data Analysis According to the information in table 1 it can be understood that the amount of T value is 7.54 with the degree of freedom (df) of 169 and in comparison with the T values it is meaningful so there are

Table 1: Comparing the average of role of information systems in information security with hypothetical average of 3 Average difference

Alpha level

Degree of freedom

T

average

0.301

0.001

169

7.54

3.3

According to the data in table 2 it can be understood that the amount of T value is 8.298 with the df of 114. This value is greater than those of T values table so it is meaningful. So there is a meaningful difference between the sample and population statistic. The amount of difference between sample average and

The role of information systems

population average is 0.492. In other words sample average of 3.49 is about 0.492 units greater than population (3) average. Therefore the role of undeniability in the security of organizational information in medical sciences organization is assessed to be average.

Table 2: Comparing the average of role of undeniability in the security of organizational information with hypothetical average of 3 Average difference

Alpha level

Degree of freedom

T

average

0.492

0.001

114

8.298

3.49


ISBN: 978-1-941968-26-0 ©2015 SDIWC

undeniabilty

population average is 0.682. In other words sample average of 3.68 is about 0.682 units greater than the population (3) average. So it can be said that the role of accessibility to information in the security of organizational information in medical sciences organization is assessed to be high.

162


Table 3: Comparing the average of role of accessibility to information in the security of organizational information with hypothetical average of 3, Average difference

Alpha level

Degree of freedom

T

average

0.682

0.001

114

14.622

3.68

According to the data in table 4 it can be understood that the amount of T value is 11.35 with the df of 114. This value is greater than those of T values table so it is meaningful. So there is a meaningful difference between the sample and

Accessibility to information

population statistic. The amount of difference between sample average and population average is 0.631. In other words sample average of 3.63 is about 0.631 units greater than the population (3) average.

Table 4: Comparing the average of role of privacy in the security of organizational information with hypothetical average of 3, Average difference

Alpha level

Degree of freedom

T

average

0.631

0.001

114

11.35

3.63

According to the data in table 5 it can be understood that the amount of T value is 11. 2 with the df of 114. This value is greater than those of T values table so it is meaningful. So there is a meaningful difference between the sample and population statistic. The amount of difference between sample average and

privacy

population average is 0.657. In other words sample average of 3.65 is about 0.657 units greater than the population (3) average. So the role of authentication in the security of organizational information in medical sciences organization is assessed to be high average.

Table 5: Comparing the average of role of authentication in the security of organizational information with hypothetical average of 3, Average difference

Alpha level

Degree of freedom

T

average

0.657

0.001

114

11.2

3.65


ISBN: 978-1-941968-26-0 ©2015 SDIWC

authentication

population average is 0.617. In other words sample average of 3.61 is about 0.617 units greater than the population (3) average. So the role of information integrity in the security of organizational information in medical sciences organization is assessed to be high average

163


Table 6: Comparing the average information integrity in the security of organizational information with hypothetical average of 3, Average difference

Alpha level

Degree of freedom

T

average

0.617

0.001

114

11.9

3.61

According to the data in table 7 it can be understood that the rating factors in enterprise information security, access to information in the first place (with an average rating of 3.25), authentication the second (with an average rating of 3.22), the

integrity

privacy of the third (with an average rating of 3.04), integration in fourth place (with average grade 2.97) and undeniably in fifth place (with an average rating of 2.51), respectively.

Table 7: Rating factors in information security Rating average

Max

min

Deviation Criterion

average

Factors

2.51

4.67

1.83

.638

3.49

undeniably

3.25

4.5

2.5

.5

3.68

Accessibility

3.04

4.86

1.71

.596

3.63

privacy

3.22

4.67

1.5

.55

3.65

authentication

2.97

4.86

2.29

.553

3.61

integration

5- Discussion and Conclusion According to the results of data analysis and hypothesis testing in the first hypothesis there is a significant difference between sample mean and population hypothetical mean so information systems have a role in the undeniability of organization information. By increasing the level of undeniability, the security of organization information increases as well. In other words in each connection neither parties can deny their cooperation. Although the advantages of information exchange is undeniable, connecting internal systems to external and international networks and delivering services and information exchange through these networks has created new risks and threats. The most important concern with regard to information systems includes accessibility of penetrators and their acts of sabotage. It is clear that in these situations physical protection methods won't

ISBN: 978-1-941968-26-0 ©2015 SDIWC

be useful so organizations have to use other methods to ensure their security. In testing the second hypothesis hypothetical average of the population has a significant difference so information systems have roles in the accessibility of organization information and by improving the level of accessibility the security is enhanced likewise i.e. information becomes available timely and for people concerned. Therefore the definition of information security policy, organization management, management of organization asset security, staff security management, physical and environmental security management organization, communications and operations security management, access to corporate information management, management of the security of maintenance and development of devices , continuity of operations, etc. should be considered.

164


In testing the third hypothesis sample average is significantly different from the hypothetical average of the population so the information systems have a role in the privacy of organization information and by increasing the level of privacy the security of organization information increases too. In other words the information is usable only by authorized people and organizations and no information is accessible to unauthorized people. Information ownership creates a context that make it possible to determine who should control the access to a particular sections of the information. Information security means protecting information and information systems from unauthorized activities. In testing the fourth hypothesis the sample average has a significant difference with the hypothetical population average so information systems have a role in the authentication of organization information and by improving the level of authentication, the security of the organization in enhanced as well. In other words both parties ensure of the authenticity of the other one. In testing the fifth hypothesis the sample average of 3.61 is significantly different from thaat of the hypothetical population so information systems have roles in the integrity of organization information and by increasing the level of integrity the security of organization information increases. In other words the information is accurate and complete. Controlling information systems especially in cases where these devices are not secure is necessary, because the dangers that may threaten these devices are too much. In companies and organizations with weak security of information systems the risk of penetration and manipulation of information to the system is high and the damage could be irreversible. The need to security necessitates that the required previsions be made by the management to ensure the suitable security of system to make the information reliable. It seems that in the above organization accessibility is more attended than other factors. Things like unauthorized modification of data, set permissions and access levels for

ISBN: 978-1-941968-26-0 ©2015 SDIWC

users, system updates and user information, delete unneeded services, etc. are more observed than other factors and the dimension of undeniability is scarcely attended. Access to information in the first place, authentication the second, the privacy of the third, integration in fourth place and undeniably in fifth place. The availability over other aspects of the organization due to be implemented should also be considered so that other aspects of the organization they are not damaged.

According to the results it is suggested that:  Proper techniques of reprimanding the staff violating the security policies be applied.  Procedures be considered for follow-up activities in the system and prevent intentional mistakes by employees.  Try to hold training courses for managers and staff both to keep them up to date and constantly remind them to keep and increase the security.  Furlough the employees who are to commit acts of sabotage.  unnecessary parts of the system that does not work be disabled  proper training courses for be considered to update the knowledge of staff  Reduce the number of employees who are responsible for information in the system and as much as possible only one person be appointed as responsible for the information system.  Robust security measures be taken and these measures be updated constantly.  To log in the system each individual use a user-defined log and staff do not disclose the username and password and it is better to use methods like fingerprints to log in.  Penetration detection systems be used and control any loggings.  Remove unused applicants  Smart filtering system intended to prevent the entry to unauthorized sites be considered.

165


 Procedures be considered in data entry to analyzed the data and prevent recording wrong data.  It should be tried to update all information regularly and all units linked to each other to make data in the whole system compatible with each other.  The entire system be backed up regularly and periodically  Proper techniques be applied for those employees violating security policies.  Procedures be considered to follow the activities in the system and prevent intentional mistakes by the employees.  Try to hold training courses for managers and staff both to keep them up to date and constantly remind them to keep and increase the security.

6.Suggestions for future research  Evaluate the relationship between the willingness of employees to use information systems and information security agency  Evaluate the role of information systems to respond to client quality  To design an information security of information systems in accordance with the structure of the organization  The challenges of using new technologies in government agencies

References [1] Taki, M., Ebrahimi, M., (1392). Information security. National Conference on Computer Engineering and Information Technology Management, Tehran, 28 June. [2] Moazzeni, Z. (1394). Information Systems Management. National Conference on Key

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Topics in Management Sciences and Accounting. [3] Gorji, A., Jafarpoor, H. (1392). Information security in the organization. Correcting and training. (134), 22-20. [4] Molavi, Z., Khenifer, H., (1393). Organizational citizenship behavior on the success of information systems in the organization. Organizational Culture Management. Volume 12, No. 1. pp. 104-83. [5] Khaksar Haghani, S., Malekshahi, A. & Khalili, M. (1392). The security of information systems. Computer Engineering and Sustainable Development Conference, Mashhad, 28 December. [6] Mehrayin, A., Ayatollahi, H., & Ahmadi, M. (1392). Information security status in the hospital information systems. Health Information Management. 10: 788-779. [7] Seifikar, M., Hamidi, N. & Hassanpoor, A. (1392). Evaluation of the effectiveness of the organization's information systems security model in different organizations. Management Conference, Challenges and Solutions, Shiraz, December. [8] Elahi, S., Taheri, M., Hassanzadeh, A. (1388). Providing a framework for human factors related to the security of information systems, Journal of human sciences teacher. (2). [9] Mahmudzadeh, A., Radrajabi, M. (1385). Iran`s Information security management systems. Journal of Management Sciences. (4): 112-78. [10] Kim S, Kim G, French A. 2015. Relationships between need-pull/technologypush and information security management and the moderating role of regulatory pressure. Information technology management. 2: 5-17. [11] Choi J, Nazareth D. 2014. A systems dynamic model forinformation security management. Information & management. 1-12. [12] Chang K, Wang CH. 2011. Information systems resources and information security. 13: 579-593.

166


Provide a Hybrid Approach to Manage Packets Motion in Order to Congestion Control in MANET Elahe Pourmazaherian1, Mohammad Reza Soltan Aghaei 2 Department Of Computer Engineering, Isfahan (Khorasgan) Branch, Islamic Azad University, Isfahan, Iran 1 [email protected] 2 [email protected] (corresponding author)

ABSTRACT Mobile ad hoc networks (MANET) is a system consists of mobile nodes that act as autonomous. The Network without any fixed infrastructure and there is no centralized management in this networks. Congestion is one of the most important challenges facing this networks, which is causing several problems. Delay and packet loss when the stations are not able to get the information because of overcrowding, are the results of congestion. In this paper, a method for congestion control is proposed that uses the AOMDV algorithm. In this method using the length of the queue of nodes, obtained routes with less traffic and basis on congestion of the founded routes, the load is distributed between them. The results of the simulation shows a decrease delay and queue length of nodes and packet loss rate and increase the throughput of the network.

increasing productivity and reducing the rate of packet loss and overhead control and energy control. If the protocols categorize based on infrastructure that used, we have the following 9 categories [1]:

KEYWORDS

- Geo Cast

MANET, Congestion Based AOMDV, Congestion Control, Load Balancing

- Power Aware

1 INTRODUCTION Routing in MANET is different from wired networks. In wired networks, the major role play by routers during routing. But there are no routers in MANET and nodes that act as a router to send packets from the source to the destination. Various protocols for MANET due to widespread using of these networks is proposed that the basic aim of all of them is

ISBN: 978-1-941968-26-0 ©2015 SDIWC

- Table driven (proactive) - On demand (reactive) - Hybrid - Geographical - Multi Path - Hierarchical - Multi Cast

In Table driven protocols, such as DSDV, nodes have a routing table that maintained the latest route information to each node in the network. Each node sends routing information periodically to their neighbors to obtain a global view of the network topology. In reactive protocols, such as AODV, unlike the proactive protocols that be updated even when do not need to send data tables, the route discovery is performed only when it needs. Hybrid protocols

167


such as ZRP, uses the specification two mentioned methods to improve them [2]. Multipath protocol in route discovery, find more than one path [3]. AOMDV algorithm is an example of this protocol. This algorithm is an improved version of AODV algorithm to find multiple paths from source to destination and then find the best path between routes as the main route and the rest are considered as alternative routes to Backup. AOMDV tries to compute multiple disjoint loop-free paths in a route discovery [3] [4]. Congestion is a situation that a lot packet exists on the part of the network. When the load on the network (the number of packets being sent through the network) more than the capacity of the network (the number of packets the network can handle), the possibility exists congestion [5]. The major criteria for monitoring congestion, the percentage of all packets discarded due to lack of buffer space, the average queue length and the number of packages that Time Out and retransmitted and packets delay average. Congestion, often occurs due to limited resources and not only lead to packet loss, long delay and decline bandwidth and wasting time operation, but also it causes a waste of time and energy in recovery time [6]. Congestion control notes techniques and mechanisms that can prevent congestion or eliminate it. The main objective of congestion control, is reduction latency and overhead buffer and also prepares network for better performance [6] [7]. 2 RELATED WORK To eliminate network congestion, many researchers have proposed the use of active queue management strategy. The basic idea is providing a buffer in order to manage or eliminate network congestion problems. Load balancing technique is another technique that

ISBN: 978-1-941968-26-0 ©2015 SDIWC

has been used in many algorithms. Another technique has been proposed that either use a way to control network congestion. In [8] RED algorithm is proposed. This mechanism is designed based on the idea that congestion detected before it happen. RED operates on the average queue size and drop packets on the basis of statistics information. If the buffer is empty, all incoming packets are authenticated. As the queue size increase the probability for discarding a packet also increase. When buffer becomes full, probability become 1 and packets are removed. In [9] when the average load of an existing link increase further the specified threshold and available bandwidth and battery power is less than the specified threshold, to reduce congestion link dense, traffic is distributed on fail-safe multiple routes. The algorithm is based on the SMORT routing protocol that computes fail-safe routes. In [10] provided AOMDV-C algorithm for load balancing. The algorithm, using the delay algorithm, calculate delay of return response. Nodes that do not respond earlier than intended, known as congested node and do not take part in communication. In This method identified the relatively idle nodes and sent by them done. [11] Proposed the ABCC algorithm that in this algorithm, all data on the network collected by mobile agents and delivered to the source. The source node selects the best route and send data. Mobile agent is a node that has a routing table that stores the routing information. In [12] an algorithm called CRP try to adaptability with congestion. In this algorithm, when one of the nodes on the main path was congested, with informing their previous node makes the previous node to seek an alternative route whit bypassing the congested nodes and

168


continue to send packets through alternative route. In [13] Congestion Based AOMDV algorithm proposed that improves AOMDV protocol. The algorithm selects the route based on the size of the queue node. Source node define some value for congestion and sends the RREQ message. Intermediate nodes with comparison their queue size with the amount that determined by the source, decide to appropriate in communications. Source node choose the best rout and rest of routes considering as alternative. 3 PROPOSED METHOD When routing is lacking congestion control, makes that in many cases chosen the wrong path by the source node. Also, always the shortest route isn’t the best choice. Choose the shortest path causes the density of nodes located in the center of the network. In the proposed method, a selection criterion is the congestion of nodes and not just its proximity to destination. The proposed method consists of two phases: route discovery and route maintenance. In the discovery phase the source node initiates routing process to send data to another node in the network. Hence prepares RREQ message to send to their neighbors. It is played by intermediate nodes in the network to obtain a route to the destination. Similar the AOMDV algorithm, to find multiple routes, all copies of the RREQ message is checked. Each intermediate node with receive RREQ packet, using the formula (1) analyzes its buffer. If the obtained value is close to a defined threshold value (the considered threshold value is 0.9), the congestion happened and node should not participate in the routing process. Otherwise, by re-send messages to their adjacent nodes, participate in communication. 𝑣𝑖 =

𝑓𝑢𝑙𝑙 𝑠𝑝𝑎𝑐𝑒 𝑜𝑓 𝑏𝑢𝑓𝑓𝑒𝑟 𝐴𝑙𝑙 𝑠𝑝𝑎𝑐𝑒 𝑜𝑓 𝑏𝑢𝑓𝑓𝑒𝑟

ISBN: 978-1-941968-26-0 ©2015 SDIWC

(1)

In this case, if the value of congestion analysis resulting close to 0.9, means that if enter more packets, a buffer overflow occurs and packets will be lost. In such a situation, a dense node by refusing to send the RREQ packet, does not participate in the routing process and at the same time, a message that there is congestion in its buffer, sends to its neighbors. If that node is a key node, which means that if this node don’t participate in routing, there is a risk of deadlock and each of the neighboring nodes that are aware of this issue, sent a message to the node, tells him who should participate in the routing process. Because if the node regardless of the participation in the routing, if a node is key, routing fail. Each node when received RREQ, added its vi to it and re-propagating. Destination receives RREQ packets, obtained average vis, using the formula (2). Then, to show the rate of change, calculated the variance with the formula (3). ∑𝑁 𝑖=1 𝑣𝑖 𝑁 ∑(𝑣𝑖 − 𝑀)2 𝑁

𝑀=

(2) (3)

Destination after calculates the mean and variance of congestion, created RREP packets and send to source node. This message is includes the calculated mean and variance of congestion at the destination. Source node with compare these values select the best path. If the mean and variance of a path is low, selected as the main route and other routes obtained as alternative routes are kept as Backup. Otherwise, using the Load Balancing, distributed load on the obtained routes. More precisely, if congestion was between 0 and 0.1, the shortest route is selected that is the most optimal route. The rest of the route kept as an alternative route when a failure occurs. If congestion was between 0.1 to 0.4, traffic distributed between two shorter obtained route and other routes are kept as an alternative route.

169


If congestion was between 0.4 to 0.7, traffic is divided between three shorter routes. The remaining routes are kept as alternative. Congestion 0.7 to 1 indicating severe traffic and load distributed on 4 shorter routes. The following figures shows sending Route request. The source node sends a message to its neighbors. During sending the nodes that are congested not sent route request message.

Figure (1-a).sending route request message

Figure (1-b).sending route reply message

In the maintenance phase when the failure occurs, RERR message is sent to the source. When the source receive a RERR message among the routes that was maintained as a backup, it selects another path with same way that choose main path. Before sending, with Hello message becomes aware of being alive the alternative path. 4 SIMULATION RESULTS OPNET is used for implementation and simulate our proposed approach and simulation is implemented in two scenarios. The first scenario where network congestion is at its worst mode. In the second scenario, the network works normally. The following parameters is considered to simulate our result.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Table1.simulation parameters

A topology of 60 nodes is established with constant bit rate traffic and the environment is an area of 2000 * 2000. Node movement is random. Routing protocol is AOMDV and transport protocol is UDP. The proposed method is compared with AOMDV and Congestion Based AOMDV algorithm. This Comparison shows that our proposed method finally did better and had a more acceptable result. When network traffic was normal case, delay the proposed approach in line with Congestion Based AOMDV and less than AOMDV, but in minute 7, our delay approach is low. In the case of heavy network traffic, our method has much less delay. Reduce the route discovery time and queue length in the proposed method, lead to reduction in overall network delay. Given that the AOMDV does not use any mechanism to control network congestion, it is natural that the more delayed than the other two methods that can be used in congestion control mechanism.

170


Figure (2-a). average delay in heavy traffic

Figure (2-b). average delay in normal traffic

Figure (3-a). average route error in heavy traffic

Figure (3-b). average route error in normal traffic

Figure (4-a). network throughput in heavy traffic

Figure (4-b). network throughput in normal traffic

ISBN: 978-1-941968-26-0 ©2015 SDIWC

171


Average route error in proposed method is less than other two methods in both normal and severe traffic. Network throughput in the proposed method compared to other methods is lower in both scenarios. Reduce the length of queues and delays and the route failure, finally leads to increase network throughput and performance.

[7]

Jung-Yoon.K , Geetam.S , Laxmi.Sh , Sarita .B, Won-Hyoung .L . 2014. Load balanced congestion adaptive routing for mobile ad hoc networks . International Journal of Distributed Sensor Networks. Vol.2014, 10 pages.

[8]

Lin, D , Morris, R. 1997. Dynamics of Random Early Detection, In Proceedings of ACM Sigcomm. Cannes, France, pp. 127-137.

[9]

Ali.M, Stewat. B, shahrabi. A , Vallavaraj. A . 2012 .Congestion adaptive multipath routing for load balancing in mobile adhoc network . International Conference on Innovations in Information Technology (IIT).

[10]

Li X , Zhi S , Xin S , Zhiyuan W, Qilong L. 2009 . Ad-hoc multipath routing protocol based on load balance and location information. In Wireless Communications & Signal Processing . International Conference on IEEE. pp. 1-4.

[11]

Sharma.V Bhadauria.s . 2012 . Mobile agent based congestion control using aodv routing protocol technique for mobile ad-hoc network . International Journal of Wireless & Mobile Networks (IJWMN) Vol. 4, pp. 299-314.

[12]

Duc.A, and Harish.R, 2006. Congestion Adaptive Routing In Mobile Ad Hoc Networks, Ieee Transactions On Parallel And Distributed Systems, Vol. 17, No. 11, pp.1294-1305.

[13]

Onkar.S , and Supratik.B, 2013. Congestion based route discovery aomdv protocol. International Journal of Computer Trends and Technology. vol.4 , No.1, pp.54-58.

5 CONCLUSIONS This paper was presented a way to congestion control. Results shows use of free congestion routes rather than shorter routes and load distribution according to network conditions and existing status and achieved routes is more appropriate. Use the load balancing one of the factors for reducing delay in the network. When there is no way out of congestion, by distributing the load can be done in less time to transfer. 6 REFERENCES [1]

Alotaibi,E, Mukherjee,B. 2012 .A survey on routing algorithms for wireless ad-hoc and mesh networks . Computer Networks 56, pp.940–96.

[2]

ZishanHaider. Y. 2013 . Performance analysis of dsdv, aodv and zrp routing protocol of manet and enhancement in zrp to improve its throughput , International Journal of Scientific and Research Publications, Vol. 3, No. 6, pp. 2250-3153.

[3]

Erfani S, Tarique M, Tepe K, Adibi S . 2009 . Survey of multipath routing protocols for mobile ad hoc networks . Journal of Network and Computer Applications 32 , pp.1125-1143.

[4]

Reddy, L. R, Raghavan, S. V. 2007. SMORT: Scalable multipath on-demand routing for mobile ad hoc networks. Ad Hoc Networks, 5(2), 162188.

[5]

Sharma,G, Akhilesh K. 2013 .Congestion control in adhoc network . International Journal of dvanced Research in Computer Science and Software Engineering. Vol.3, No.6, pp.283-286

[6]

Senthilkumaran.T, Sankaranarayanan.V . 2012 . Dynamic congestion detection and control routing in ad hoc networks . Journal of King Saud University Computer and Information Sciences,pp. 25-34.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

172


Improvement of Load Distribution and Prevention of Congestion in Ad-Hoc Wireless Networks Using Classification and Swarm Intelligence Somaye Mobini1 & Mohammad Reza Soltan Aghaei2 1. Graduating student of computer Engineering, Islamic Azad University, Isfahan (Khorasgan) Branch, Isfahan, Iran. Email: [email protected] 2. Assistant professor of computer group, Islamic Azad University, Isfahan (Khorasgan) Branch, Isfahan, Iran. Email: [email protected] (Coresponding)

ABSTRACT Mobile ad hoc networks consist of mobile platforms which are formed dynamically by a group of mobile hosts. The hosts have a capability to detect, navigate and recover the link automatically. Creation of a highly reliable routing is one of the most important factors in networks. Load balance and network congestion are major issues in creation of routing. Disproportionate load distribution in networks causes network performance to reduce considerably. In the proposed method, the ant colony optimization algorithm searches high quality routes using agents. Then, using data classification and balanced load distribution, multiple routes are used instead of leading all the traffic on one singe route with an aim at prevention of congestion in the network and reduction of delays. The results indicate that in addition to reduction of delays, performance of the networks is also enhanced.

KEYWORDS data classification, multiple routes, balanced load distribution.

1 INTRODUCTION The mobile ad hoc network is a temporary wireless network that is formed dynamically by a group of mobile hosts. Routing protocols of ad hoc networks should be efficient, have low energy consumption and low computation consumption and time [1]. Simplicity and speed of network setup is one of the features of mobile ad hoc networks. Route maintenance is periodically carried out to possess the optimum

ISBN: 978-1-941968-26-0 ©2015 SDIWC

route. This operation is accomplished by data packets. Ad hoc routing protocols are divided into two types of active and reactive. Active routing protocols are based on the maintenance of routing data compatibility in each node by its propaganda across the network. Reactive protocols find a route from a source to a destination only when routing is requested by the source. Creation of a highly reliable routing is one of the most important factors in networks. Load balance and network congestion are major issues in creation of routing. Disproportionate load distribution in networks causes network performance to reduce considerably. The use of data classification mechanisms and balanced load distribution causes the network performance to enhance. Using leading and backward agents, the ant colony optimization algorithm searches high quality routes and instead of leading the traffic on a single route releases the traffic on multiple routes to prevent network congestion and increase the reliability of delivery of packets to destination. In this paper, packets are classified in accordance with their service type and based on the requirement of packets, they are allocated a route and the load is distributed on the routes. For example, packets containing a video stream or those with a volume of above x, in which bandwidth and delay are important, are given four routes, while packets containing audio, in which delay is important, are given three routes and in the same way, routes are assigned to

173


packets and the load is distributed on the routes. In principle, a route is assigned in accordance with the network traffic. With a balanced load distribution on routes according to packets traffic, it is possible to reduce the network delay and packet retransmission and also increase network efficiency. 2 PREVIOUS WORKS According to the conducted research, some algorithms were investigated from the perspectives of type (active, reactive or combinatory), the number of search routes (single or multiple), use or non-use of ants to find optimum routes, method of sending ants (periodic, regular intervals and more) and also the way of using backup routes (alternative route in a table, discovering the rout on demand and more). Algorithms MABR [4, 10], PERA [4, 11], AMQRA [4, 19], Ant Net [2], DSDV [6], WRP [6, 12] and ABC [3, 5, 9] are active algorithms that perform their routing based on a routing table and update the tables when the network topology changes. ARA [14], ADRA [5], MANSI [5], DSR [7], AMQR [5, 19], AODV [6, 13], PACONET [4, 5, 11], PBANT [4, 8] and Ant-E [4, 16] were introduced as reactive algorithms that search the routes based on the demand to transfer the traffic, while AntAODV [4, 22], ARAMA [4, 17], HOPNET [4, 5], Ant-HocNET [15] and ZRP [6, 14] are combinatory algorithms that present higher scalability than the other two algorithms. Algorithms MABR, ARA, MANSI, ADRA, ABC, Ant-AODV and HOPNET create exploring ants periodically to discover optimum routes in a network. Algorithms PERA, AMQRA, Ant Net, PACONET and PBANT send ants into the network at regular intervals to discover the needed routes. Algorithms MABR, PERA, AMQRA, Ant Net, DSDV, WRP, AMQR, ABC, PACONET, AntE and ZRP use a routing table to substitute a backup route, but algorithms ARA, ADRA, MANSI, AODI and AntHocNet discover a backup route based on the demand.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Algorithms MABR, PERA, Ant Net, DSDV, WRP, ADRA, AODV, PACONET, AntAODV, ARAMA and ZRP use a single route to transfer the traffic, while algorithms AMQRA, ARA, MANSI, DSR, AMQR, Ant-E, HOPNET, FDAR [3, 21], FSMR [4, 18], AOMDV [13, 20] and AntHocNet use several routes to transfer the traffic. In this study, the algorithm Ant-AODV and the research results of references [1, 4, 22] were used. Algorithm Ant-AODV, which is a combination of ants’ behavior and basic AODV, aims at overcoming the inherent shortcoming of AODV in which a constant number of independent ants move in a network randomly and maintain the nodes last visited and update the routing tables actively. The algorithm reduces the end to end delay and increase the number of links between nodes. Therefore route discovery is conducted more rapidly. Finally, ant agents update the routes continuously. Thus, the source node can be changes from an old route to a new shorter route which is presented by ants. As a result, it leads to a significant reduction in the end to end delay than algorithms of AODV and ants based routing. The algorithm, like AODV, notifies upstream nodes of the local links failure by sending route failure messages. The two algorithms table (AODV and ants based routing) implements the routing table in AntAODV and in order to keep the neighboring table of repeated broadcast, HELLO is used if the node had not already been visited by ants. In the carried out simulation, the packet overhead as well as the end to end delay is less than AODV while the amount of received packets is at the same level of AODV. In fact, it can be said that in this idea, along with a change in links position, leading ant are generated and move in the network [4, 22]. In a paper entitled “balanced load distribution and guarantee of service quality in ad hoc networks using swarm intelligence agents”, ant agents are used to discover routes. In addition, traffic classification is mentioned in which traffics are put in distinguished classes based on

174


a need to bandwidth and the delay and in order to determine the service class, two bits are added to the exploring ants’ packet. In this paper, two routes are used to transfer the traffic in which the optimum route is assigned to the higher service class traffic and the auxiliary route is dedicated to the lower service class traffic. This service classification causes links traffic and delay to be reduced. It will also play an effective role in the balance distribution of load in the network [1]. 3 PROPOSED METHOD Since the topology of ad hoc networks is constantly changing and radio interference causes many packets to be lost, guaranteeing reliability in these networks is very difficult. In a group of protocols, it is possible to send several copies of data on several separate routes to enhance reliability and tolerance against failures. Although this method increases power consumption, data loss probability decreases in case of link failures. In this study, exploring ants discover multiple separate routes from a source to a destination and simultaneously distributing the traffic on the discovered routes result in the load balance as well as the increase of network efficiency. As seen in figure (1), exploring ants move from the origin to the destination in accordance with the proposed algorithm flowchart and discover N existing routes. If the number of discovered routes is four or higher, four of the higher quality routes are selected and if they are less than four routes, all of them are chosen and assigned to data packets based on their service class. For example, if the packet service class is (00), it is distributed among four load routes and then the packets are sent along the optimum routes. If a route is broken, route discovery starts from the beginning.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Figure (1): Flowchart of the proposed algorithm

4 SIMULATION METHOD

of

the

PROPOSED

The proposed algorithm was simulated along with Ant-AODV and the data classification based algorithm. Three scenarios were run including limited, average and high volume traffic injection. Then, the results obtained from the proposed algorithm as well as the other two algorithms in the above mentioned scenarios were assessed. Owing to the high number of diagrams, the results obtained from two scenarios (limited and average volume traffic injection) were ignored in this paper and were only mentioned in the results. However, the

175


diagrams of the third scenario, that is high volume traffic injection to network, are presented. In the diagrams obtained from the results, green, red and blue colors respectively denote the proposed algorithm, the data classification algorithm and Ant-AODV algorithm In this scenario, all three algorithms are injected the same amount of traffic. As seen in figure (2), all three protocols exhibit the same amount of traffic because they are in the same condition.

Figure (3): The amount of received bits

As seen in figure (4), the amount of deleted packets in the proposed algorithm has been reduced considerably in comparison with the data classification based and Ant-AODV algorithm.

Figure (2): Injection of traffic to the network

As seen in figure (3), the amount of received bits is of the same size in all three protocols up to about 40000 bits. However, from 40000 upwards, the received bits of the proposed algorithm are significantly higher than the data classification based and Ant-AODV algorithm.

Figure (4): Deleted packets

As seen in figure (5), the average retransmitted packets have been considerably reduced in comparison with the Ant-AODV and the data classification based algorithm.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

176


Figure (5): Retransmission of packets

Figure (7): The time needed to discover a route

According to figure (6), the average traffic created for routing in the network in the proposed algorithm has been reduced with respect to Ant-AODV and the data classification based algorithm.

In the proposed algorithm, the average network delay time has been reduced considerably in comparison with Ant-AODV and the data classification based algorithm as shown in figure (8).

Figure (6): The traffic created by protocols to discover a route in the network

Figure (8): Network delay

The average time needed to find a route in the proposed algorithm shows a considerable decrease in figure (7). Injecting a high volume of traffic, the data classification based and AntAODV algorithms have totally lost their balance for finding a route.

According to figure (9), in the proposed algorithm (balanced load distribution together with data classification), injecting a high volume of traffic to the network, the amount of bits per second lost as a result of buffer overflow is close to zero, while in the two other algorithms, as the injected traffic volume

ISBN: 978-1-941968-26-0 ©2015 SDIWC

177


increases, more bits are lost per second as a result of buffer overflow.

Figure (9): The data lost as a result of completion of buffer capacity

As seen in figure (10), the high volume traffic injection to the network causes the average efficiency of the proposed algorithm to be increased considerably in comparison with algorithm Ant-AODV and the data classification based algorithm.

Figure (10): Average efficiency

Generally, simulating the proposed algorithm as well as algorithm Ant-AODV and the data classification based algorithm, running three traffic injection scenarios of limited, average

ISBN: 978-1-941968-26-0 ©2015 SDIWC

and high volumes and also evaluating the behavior of the three algorithms, some results were obtained as follows: Running the first scenario which is the limited volume traffic injection, packet elimination in the proposed algorithm is lower than AntAODV algorithm and the data classification based algorithm respectively by 37.75% and 27.85%. Running the second scenario which is the average volume traffic injection, packet elimination in the proposed algorithm is lower than Ant-AODV algorithm and the data classification based algorithm respectively by 37.99% and 19.01%. Running the third scenario which is the high volume traffic injection, packet elimination in the proposed algorithm is lower than AntAODV algorithm and the data classification based algorithm respectively by 154.89% and 144.31%. In the first scenario, the delay of the proposed algorithm is at the same level as Ant-AODV and the data classification based algorithm. In the second scenario, the delay of the proposed algorithm is at the same level as AntAODV and the data classification based algorithm. In the third scenario, the delay of the proposed algorithm is zero and lower than Ant-AODV and the data classification based algorithm respectively by 6425% and 7625%. In the first scenario, the proposed algorithm traffic is equal to Ant-AODV algorithm and lower than the data classification based algorithm by 8.79%. In the second scenario, the proposed algorithm traffic is the same as Ant-AODV and the data classification based algorithms. In the third scenario, the proposed algorithm traffic is lower than Ant-AODV and the data classification based algorithm respectively by 14.73% and 15.91%. Running the first scenario, the operational power of the proposed algorithm is higher than Ant-AODV algorithm by 1.8%, while it is

178


lower than the data classification based algorithm by 6.1%. Running the second scenario, the operational power of the proposed algorithm is higher than Ant-AODV algorithm and the data classification based algorithm respectively by 6.09% and 1.3%. Running the third scenario, the operational power of the proposed algorithm is higher than Ant-AODV algorithm and the data classification based algorithm respectively by 35.82% and 39.96%.

[2]

Di Caro, G. and M .Dorigo (2011). "AntNet: Distributed stigmergetic control for communications networks." arXiv preprint arXiv:1105.5449.

[3]

Gupta, A. K., H. Sadawarti, and A. K. Verma (2012). "Computation of Pheromone Values in AntNet Algorithm." International Journal of Computer Network and Information Security )IJCNIS) 4(9): 47.

[4]

Wankhade, S. and M. Ali (2012). "Recent trends in ant based routing protocols for manet".

[5]

Lutimath , N. M., D. G. Anand ,and L. Suresh (2012). "A Survey of Ant based Routing Algorithms for Mobile Ad- hoc Network." International Journal of Advanced Research in Computer Science and Software Engineering 2(8): 89-93.

[6]

Talwar, B. and A. K. Gupta (2012). "Ant Colony based and Mobile Ad Hoc Networks Routing Protocols: a Review." International Journal of Computer Applications 49(21): 36-42.

[7]

Johnson, D. B., D. A. Maltz, et al. (2001). "DSR: The dynamic source routing protocol for multi-hop wireless ad hoc networks." Ad hoc networking 5: 139-172.

[8]

Sujatha, B. and D. Sathyanarayana (2010). "PBANT-Optimized ANT Colony Routing Algorithm For Manets." Global Journal of Computer Science and Technology 10(3).

[9]

Schoonderwoerd, R., O. Holland, and J. Bruten (1997) .Ant-like agents for load balancing in telecommunications networks. Proceedings of the first international conference on Autonomous agents, ACM.

[10]

Heissenbüttel, M. and T. Braun (2003). Ants-Based Routing in Large Scale Mobile Ad-Hoc Networks. KiVS Kurzbeiträge.

[11]

Osagie, E., P. Thulasiraman, and R. K. Thulasiram (2008). PACONET: imProved ant colony optimization routing algorithm for mobile ad hoc networks. Advanced Information Networking and Applications, 2008. AINA 2008. 22nd International Conference on, IEEE.

[12]

Murthy, S. and J. J. Garcia-Luna-Aceves (1996). "An efficient routing protocol for wireless networks." Mobile Networks and Applications 1(2): 183-197.

[13]

Abolhasan, M., T. Wysocki, et al. (2004). "A review of routing protocols for mobile ad hoc networks." Ad hoc networks 2(1): 1-22.

[14]

Gupta, A. K., A. K. Verma, and H. Sadawarti (2011). Analysis of various Swarm-based & Antbased Algorithms. Proceedings of the International Conference on Advances in Computing and Artificial Intelligence, ACM.

5 CONCLUSION Simulating the proposed algorithm, algorithm Ant-AODV and the data classification based algorithm and running three scenarios of limited, average and high volume traffic injection, the results obtained indicate that the proposed algorithm reduces network delays, route search delay and packet elimination in the network because of different reasons such as buffer overflow, loss of packets and also retransmission of packets in the network and also presents a desirable operational power. Its advantage compared to algorithm Ant-AODV and the data classification based algorithm is that its stability when facing high volume traffic is high. Finally, given the improvements resulting from the proposed algorithm compared with AntAODV and the data classification based algorithm, it is suggested that the protocol is evaluated for other wireless networks such as sensor networks, vehicles wireless networks (VANET) and also it is implemented along with security methods so that the algorithm becomes more efficient. REFERENCES [1]

Khodami, A., M, Soltan Aghaei (2013). Balanced load distribution and service quality assurance in Ad-hoc network using swarm intelligence agents. The second Symposium of engineering equipment Contacts[In Persian].

ISBN: 978-1-941968-26-0 ©2015 SDIWC

179


[15]

Di Caro, G., F. Ducatelle, and L. M. Gambardella (2004). AntHocNet: an ant-based hybrid routing algorithm for mobile ad hoc networks. Parallel Problem Solving from Nature-PPSN VIII, Springer.

[16]

Sethi, S. and S. K. Udgata (2010). "The efficient ant routing protocol for MANET." International Journal on Computer Science and Engineering 2(07): 24142420.

[17]

Hussein, O. and T. Saadawi (2003). Ant routing algorithm for mobile ad-hoc networks (ARAMA). Performance, Computing, and Communications Conference, 2003. Conference Proceedings of the 2003 IEEE International, IEEE.

[18]

Dharaskar, R. V. and M. Goswami (2009). "Intelligent Multipath Routing Protocol for Mobile AdHoc Network." International Journal of Computer Science and Applications 2(2): 135-145.

[19]

Umlauft, M. and W. Elmenreich (2008). QoS-aware ant routing with colored pheromones in wireless mesh networks. Proceedings of the 2nd International Conference on Autonomic Computing and Communication Systems, ICST (Institute for Computer Sciences, Social-Informatics and Telecommunications Engineering).

[20]

Yuan, Y., H. Chen, and M. Jia (2005). An optimized ad-hoc on-demand multipath distance vector (AOMDV) routing protocol. Communications, 2005 Asia-Pacific Conference on, IEEE.

[21]

Wang, X., S .Tagashira, et al. (2007). "FDAR: A Load-Balanced Routing Scheme for Mobile Ad-Hoc Networks." 4686: 186-197.

[22]

Marwaha, S., C. K. Tham, and D. SRINIVASAN (2002). Mobile agents based routing protocol for mobile ad hoc networks. Global Telecommunications Conference, 2002. GLOBECOM'02. IEEE, IEEE.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

180


Platforms for Use Integrated Resources Formative Processes in E-learning P. Campanella Department of Computer Science University of Bari “Aldo Moro” via Orabona,4 – 70126 Bari – Italy [email protected] ABSTRACT In recent years there has been a rapid development of platforms for distance learning. In this paper we study solution open source with particular reference to widespread, typical of web 2.0, namely feedback, chat, blog, forum, podcasting, wiki, facebook, youtube and skype, in order to better manage online courses interactive, which makes the user does not only more user content but even an active participant in the process production; in this way the system becomes a tool for sharing knowledge.

analysis of some of the most popular platforms open-source that wants to be a useful contribution with reference to the development of different forms of collaborative learning, typical web 2.0 and require new skills in order to integrated management of training components typical of social networks. Following the different sections on platforms open source analyzed in their studies and finally conclusions and future developments.

KEYWORDS

2 OPEN SOURCE PLATFORMS platform; e-learning; tools; web 2.0; report

1 INTRODUCTION In field of e-learning activities that allows the protagonist is platform, which is set of all those technological supports that allow management of an course online, integrating products educational, testing and evaluation and exchange interactions within learning groups prepared [2,3]. But the tools of the different platforms may have different features between them: simple delivery of content but also complexity functional, technology and education, just to the flexibility of groups than in the form of information retrieval [8,10,12]. At the base of an platform e-learning there is therefore a modular architecture, whose characteristic is that of being able to aggregate modules of the same level without altering mode of operation of existing elements. In an effort to encourage the use of more and more platforms customizable it has proposed modular

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Table that follows reports different open source platforms exist today which have characteristic of having a wider freedom of action to a minor implementation cost compared to those commercial specifying contents present in them in respect to sharing, collaboration or (Table 1): blog, feedback, forum, podcasting, wiki [5,6,8,12]. Table1. OPEN SOURCE PLATFORMS

Innovation brought integrates typical tools of web 2.0 or facebook, youtube and twitter in order to better facilitate collaboration between users, thus offering great flexibility to the system.

181


ADA LMS constructivist acronym Digital Environment for Learning, developed by Lynx, and multimodal, allowing easy sharing of structured content into pedagogical hypertext and hypermedia and interactive cooperation between users [5,6,9]. The experimentation conducted involved the distribution of a questionnaire with the purpose of having an evaluation has the courses that the instrument used by teachers. An analysis of the answers it was noticed that (Figure 1) in general the duration of the course was short and calls for greater usability of the platform over time. The material made available on the network was found to be good or very good, with percentages of 57% and 43%, as well as the quality of lectures, exercises and tests which was evaluated average, with an average of 49,5%. For which the monitoring was on average balanced.

Figure 2. Report integrated modules ADA

Atutor LCMS platform founded in 2002 as a result of a collaborative project of ATRC (Adaptive Technology Resource Centre), University of Toronto. It supports the IMS/SCORM in terms of reusability [9,12]. Tools atutor can be classified into communication tools, productivity, support for the administrator, distribution courses. Modular, multi-platform, supports systems windows, linux, mac os. Based on MySQL/ PHP/Apache [5,6,7,8].

Figure 1. Monitoring ADA

The experimental test administered to 100 users in the community gave the following answers about the three forms youtube, facebook, skype; a challenge to be improving at using the tools after consultation of the courses currently estimated at two hours, in general the degree of user satisfaction in the use is shown in Figure 2, to youtube and facebook is good, with a percentage of 32% and 35%, for skype it is average. Monitoring tests conducted on average balanced.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Figure 3. Screen integrated modules in Atutor

The features in this platform are quite useful for advanced customizing the interface [5,6,9]. The experimental test has been to administer a questionnaire to 200 users in the community who gave the following answers on the two modules facebook, skype integration (Figure 3). A challenge in the process of improvement will be to make the last use of tools after

182


consultation of the courses, in general the degree of user satisfaction in the use is shown in the graph below (Figure 4), for facebook it is very good, with a percentage of 55% and for skype is good, with a percentage of 40%.

Figure 4. Report integrated modules in ATutor

Claroline LCMS platform of e-working, developed by the University of Louvain (Belgium). Based on PHP/MySQL. Its graphics facilitates students in navigating and finding resources [13]. Modular platform for windows, linux and mac os [5,6,9]. The experimentation conducted consisted of the distribution of a questionnaire to 100 users in the community who gave the following answers on the two ad hoc modules facebook, skype integrated. A challenge in the process of improvement will be to make the last use of tools after consultation of the courses, in general the degree of user satisfaction in the use is shown in the graph below (Figure 5), for facebook it is very good, with a percentage of 60% and for skype is good with a percentage of 40%. Monitoring test conducted was quite optimal.

Figure 5. Report forms in Claroline

Docebo LMS was born in Italy, initially Spaghetti Learning, highly customizable [11,13]. It allows you to provide training courses in different teaching methods: self-learning, blended learning, collaborative, and sociallearning. The testing administered to 100 users in the community reported a higher average degree of satisfaction in using even about the ad hoc modules to meet specific needs, or youtube, facebook, skype. Monitoring conducted on average balanced. EifFel LMS high scalability, accessibility and portability [3,9]. It allows you to provide training courses in social learning mode. Supports SCORM. The experimentation conducted shows a graphical representation of the log file to evaluate the user-system interaction concerning access platform in the

ISBN: 978-1-941968-26-0 ©2015 SDIWC

183


period from 2 January 2013 to 16 July 2013. A reading of the graph (Figure 7) shows, with reduced variances and exceptions, a constant trend and generalized of attendance by users, the inhomogeneity is had for the lack of customization of the platform.

Figure 7. Graphic frequency users access Eiffel

Ilias Groupware platform, Integrated Information and Learning Co-operative Work System, developed by the University of Cologne in Germany. Cooperation allows synchronous and asynchronous, modularity, flexibility and customization [13]. Multiplatform systems windows, linux, mac os. Developed in PHP, it supports IMS, AICC, SCORM, LDAP authentication [14,15]. The experimentation conducted has led to an improvement in the management of the menus, instead monitor (Figure 8), tracking was a little complicated as evidenced by the graphical login, in fact they have been characterized more by exploration activity that has led to investigate the resources and tools available, it is evident jumps and changes of direction.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Figure 8. Launch course Operating Systems and tracking users ILIAS

Moodle LCMS platform, based on the pedagogical principle of social constructivism, an acronym for Modular Object-Oriented Dynamic Learning Environment [1,14]. Developed by Martin Dougiamas, at Curtin University, Australia, in PHP. And it is written in 34 languages. Represents the evolution of the old systems VLE very close to the Personal Learning Environment (PLE) [5,6]. The platform comes with a simple, lightweight and usable. Supports SCORM/AICC [4]. The experiment conducted involved at the community level in total 60 teachers who were administered a semi-structured interview. Following a report of the activities carried out by members, progress, attendance consultation materials and level of participation in collaborative activities as well as an assessment of the ad hoc modules created in the platform to meet specific needs or facebook, skype,

184


youtube and twitter (Figure 9) showing a very good for facebook and skype, with a percentage of 50% and 45%, a good for youtube with a percentage of 45% and an average for twitter, with a percentage of 34% (see Figure 10). Monitoring users optimal experimentation started.

Figure 9. New management modules integrated in Moodle 2.0

3 CONCLUSIONS AND FUTURE DEVELOPMENTS In conclusion in this paper we were considered open source solution in order to carry out a study of targeted analysis, or the identification of those features considered essential considering the continuous evolution that tends to place the user at the center in the use of new content. In this scenario was considered the new web 2.0 leading to the spread of the triad collaboration, participation and sharing, and were integrated those features typical of the new social web or feedback, chat, blog, forum, podcasting, wiki, facebook, youtube and skype not present but that lead the user to interact at the level of community, transforming the system from a simple box of teaching materials in tool for sharing and knowledge management. The experiments that particularly concern the self-study, asynchronous collaboration through forum or discussions and synchronous with direct communication and sharing applications have progressed well considering the traceability index and the results of questionnaires administered to students, in terms of satisfaction, acquisition of knowledge and changes in performance. There are still improvements that hit the interface on most platforms is not always the type user-friendly and the lack of flexibility found in some platforms analyzed. REFERENCES [1] M. Acquaviva, “Learning management systems, Open Source a confronto”, iGeL, Il Giornale dell’elearning, anno 1, n. 2, 2013. [2] A. Andronico, A. Chianese, B. Ladini, “E-Learning: metodi, strumenti ed esperienze a confronto”, Didamatica 2002, Liguori Editore, Napoli, Italia, 2002. [3] M. Banzato, D. Corcione, “Piattaforme per la didattica in rete”, TD-Tecnologie Didattiche n. 33, pp. 22-31, 2004.

Figure 10. Management Tracking user access Moodle 2.0

ISBN: 978-1-941968-26-0 ©2015 SDIWC

[4] Beccacene P., “E-learning: la scelta di un Learning Management System open source e la creazione di pacchetti SCORM”, In AlmaTwo, Osservatorio elearning, 2005.

185


[5] P. Campanella, “Learning Management Systems: A comparative analysis of open-source and proprietary platforms”, in Proceedings of IADIS International Conference e-Learning 2011, 20-21-22-23/07/2011, Roma, Italy, ISBN: 9789728939526.

Learning 2011, 20-21-22-23/07/2011, Roma, Italy, ISBN: 9789728939526.

[6] P. Campanella, “Functional Comparison of the Tools and Commercial Platforms in Distance E-Learning”, in Proceedings of IADIS International Conference eLearning 2011, 20-21-22-23/07/2011, Roma, Italy, ISBN: 9789728939526. [7] P. Campanella, “Method of experimental evaluation of ICT in teaching”, Atti del convegno Elearn 2011, World conference on E-learning in Corporate Governement, Healthcare e Higher Education organized by AACE, 17-18-19-20-21 October, 2011, Honolulu, Hawaii, USA. [8] P. Campanella, Piattaforme per l’Uso Integrato di Risorse Formative nei Processi di e-learning, Atti DIDAMATICA 2015 – Studio Ergo Lavoro – dalla società della conoscenza alla società delle competenze, 15-16-17/04/2015, Genova, Italy, ISBN: 978-88-98091-38-6. [9] P. Campanella, Piattaforme proprietarie: Un’analisi metodologica, Atti DIDAMATICA 2015 – Studio Ergo Lavoro – dalla società della conoscenza alla società delle competenze, 15-16-17/04/2015, Genova, Italy, ISBN: 978-88-98091-38-6. [10] P. Campanella et al, Content Management System Open Source: Un’analisi comparativa, Atti DIDAMATICA 2011, Informatica per la didattica, 04-05-06/05/2011, Torino, Italy, ISBN: 9788890540622. [11] P. Campanella, NetLearn2.0: Piattaforma e-learning e metodologie integrative, Atti Didamatica 2013, Tecnologie e Metodi per la Didattica del Futuro, 0708-09/05/2013, Pisa, ISBN: 978-88-98091-10-2. [12] P. Farace, “Strategie nell’e-learning: l’impatto del modello open source nelle scelte tecnologiche e funzionali”, 2003. [13] S. Impedovo, P. Campanella, G. Facchini, G. Pirlo, R. Modugno, L. Sarcinella, “Learning Management Systems: un’analisi comparativa delle piattaforme open-source e proprietarie”, DIDAMATICA 2011 Informatica per la didattica, 04-05-06/05/2011, Torino, Italy, ISBN: 9788890540622. [14] S. Impedovo, IAPR Fellow, IEEE S. M., P. Campanella, “LMS: Benchmarking ATutor, Moodle and Docebo”, IADIS International Conference eSociety 2012, 10-11-12-13/03/2012, Berlin, Germany, ISBN: 978-972- 8939-66-3. [15] S. Impedovo, P. Campanella, G. Facchini, G. Pirlo, R. Modugno, L. Sarcinella, “A Comparative Assessment of e-Learning Platforms”, in Proceedings of IADIS International Conference e-

ISBN: 978-1-941968-26-0 ©2015 SDIWC

186


Learning Management Systems: A Comparative Analysis of Open-Source and Proprietary Platforms P. Campanella Department of Computer Science University of Bari "Aldo Moro" Via Orabona, 4 – 70126 – Bari (Italy) [email protected] ABSTRACT In the context of distance education, the choice of platform for content delivery becomes a critical success factor. This paper analyzes the state of the art of e-learning platforms, Learning Management Systems and sums up a comparative analysis of some of the most popular open-source and proprietary. I test results on the comparison between the 27 different platforms and e-learning, show that at present there is still a significant gap between open source and proprietary platforms, particularly with regard to the characteristics of social networks.

KEYWORDS synchronous distance e-learning; multimedia tools; e-learning tools; functional evaluation; collaborative learning; e-cooperation

1 INTRODUCTION In recent years, the use of e-learning technologies is particularly increased, making possible new ways of learning-based software platforms, today more and more customizable [7,9,10]. These technologies from those of the web 2.0, define the nature of the new e-learning 2.0, which also includes those non-technical aspects that are combined with the inherently social nature of the network [4,12,19,20,42] and allows interaction with learners in a variety of innovative services for collaborative learning [12,17,33,36]. Much has been made in consideration that most modern theories prefer informal learning at the expense of the old conceptions [2,14,21,25,43]. In particular, according to [19], e-learning 2.0 should

ISBN: 978-1-941968-26-0 ©2015 SDIWC

integrate all these new learning techniques typical of web 2.0 tools to allow users an effective dialogue on the Internet [29,34,35,40,42]. With typical tools of web 2.0, it is therefore possible for teachers and students to share content through blogs, podcast, social bookmarking also by the addition of words with keys link (tagging) in an interactive and flexible that create real virtual learning environments [6,22,42,3,21,23,25,36,39]. In this scenario, despite several reported in the literature, it is still difficult to reach and shared evaluation of effective e-learning platforms, which have undergone a remarkable evolution in recent years exactly with the development of web 2.0, considering the different application domains of use [30,31,38,40,69]. This paper first presents the characteristics of open source and proprietary platforms for analysis, followed by a comparison of them, especially 10 and 17 commercial open source on the presence of features typical of the so-called web 2.0 “social networks”, as well as discussion of the results and conclusions. 2 PLATFORMS OPEN-SOURCE AND PROPRIETARY The e-learning platforms are divided into two broad categories: open-source and proprietary [8,11,27]. Here is a table that assessment for the various parameters considered very important or license, market control, innovation, safety, cost, scalability in terms of cost, media platforms provides results about their business

187


and technology management [32,49,56,67,68]. It should be noted that the open-source turn out to be more balanced than the more restrictive proprietary. TABLE 1. PLATFORMS OPEN-SOURCE TO PLATFORMS PROPRIETARY

online, allow individual participants to be fully interactive [1,25]. These three elements, their interrelations with each other, create the paradigm of “Collaborative Learning”, which combines the technological and organizational aspects of the learning process [24,26,41,43,70]. 3 PLATFORMS E-LEARNING: A COMPARATIVE ASSESSMENT

The Table 1 then provides the basis for understanding the current phenomenon of the gradual, relentless training by static characteristics typical of proprietary platforms to the typical dynamics of open source, which walks you through the perspective of lifelong learning so desired [5,21,25,41]. Faced with the prospect of growing e-learning, it is necessary to question the main factors that determine the success [15,28,29]. In particular, the efforts of researchers in the field are moving more and more identified with the factor of collaboration: ° the use of Open Source technologies (OS); ° Peer (P2P); peer-to-Peer systems (P2P); ° Creating Virtual Learning Community (VLC) for learning. OS technologies are, by definition, the world that creates value in a collaborative way, customizing the online learning systems. P2P systems, the first example of collaboration between computers, allowing you to easily expand the network of participants by offering greater availability of resources. Finally, the VLC, the new models that represent the learning support intra-and inter-organizational

ISBN: 978-1-941968-26-0 ©2015 SDIWC

This section presents a comparative assessment of the main platforms for e-learning, with particular reference to its ability to integrate the most popular services, which are typical of web 2.0, “social network”, like facebook, feedback, chat, blog, forum, youtube, podcasting, skype and wiki [4,13,18,19,23]. Table 2 shows in particular the analysis of ten leading platform for open-source ada, atutor, claroline, docebo, eiffe-l, freelearn, ilias, moodle2, .LRN, sakai [44,45,46,47,48,49,50,51,52,53,54,55,56,57,58, 59,60,61,62,63,64,65,66,67,68,69]. It is noted that some services such as feedback, blog, forum, podcast and wiki are integrated across all platforms, but no open-source platform supports skype and youtube. Instead, facebook is only supported by the platforms docebo, moodle2 and .LRN, while the platform .LRN is the only one not supporting the chat. TABLE 2. PLATFORM FEATURES OPEN-SOURCE

Table 3 shows the analysis, instead of seventeen proprietary platforms, including the most common: adobe connect, centra,

188


elluminate live, e/pop, groove, hotconference, megameeting, netlearning, picturetalk, raindance, saba learning enterprise, same time, t-learn, voxware, webct, wave three, webconference. It follows in particular that all platforms considered complement tools such as feedback, chat, blog, forum, podcasting and wiki, but none supports facebook, youtube and skype. TABLE 3. FEATURES PROPRIETARY PLATFORM

The results reported in Tables 2 and 3 show that at least some products, there is a substantial difference between the different platforms are analyzed as the sweeping changes favored by the free market tend to do so it is equipped with all the typical instruments that make them comparable. Of particular importance is the fact that the tools more widely available, such as youtube and skype are not integrated into any of the platforms examined both open-source and proprietary [13,16,37,71]. In open source platforms is also more visible instead of the use of tools to support the development of social networks, typical of web 2.0 [36]. 4 DISCUSSION AND CONCLUSIONS This paper is the consideration that the elearning market is expanding as is the growing need for new methods and training tools for new users. In particular, the actual differences

ISBN: 978-1-941968-26-0 ©2015 SDIWC

between the different solutions is very important to identify those that best meet their training needs. The analysis of the platforms analyzed in this article is intended as a useful contribution to the evaluation of the same in relation to the development of various forms of collaborative learning and the needs of web 2.0, and require new skills on the integrated management of multiple educational components, typical of social networks. The results show that, given the large number of possible solutions on the market, only very few platforms integrate some typical instruments of social networks. In some cases, the structural flexibility in defining the content, enabling the platform to adapt to their training needs, for example through the use of modules or plug-in. In most cases, however, the analysis of the peculiarities of the platforms, both proprietary and open-source, demonstrates that the rigidity and lack of flexibility are characteristics common to many of them. This is particularly true in the field of proprietary platforms, which are also facing the apparent presence of several products on the market actually offered is rather limited, since all the solutions are similar if not identical. In conclusion, it can be argued that the analysis highlights the extreme rigidity of elearning platforms currently available and their general lack of attention, less than a few isolated exceptions, many important aspects of social networking. Currently he is also evident is a significant gap between some open-source platforms, particularly advanced for their support in order to update the community and to enhance their skills, late in the features typical of web 2.0. The progress of research and the commitment of the developers is therefore still needed in this area, where, however, important new requirements are emerging with the use of platforms more modern forms of ebusiness, coaching and development of peer-to -peer. In this scenario must also be considered that the development of technologies and systems, changes in content and in the development of new tools can not be done without considering the needs of the user are

189


becoming more sophisticated, since the training does not longer limited to the initial location, but continues throughout life (lifelong learning). REFERENCES [1] R. C. Atkinson, R. M. Shiffrin, “A proposed System and its Control Process”, In Psychological of Learning and Motivation: Advances in Research and Theory, vol.2, pp.89-105, 1968, New York Academic press.

Ergo Lavoro – dalla società della conoscenza alla società delle competenze, 15-16-17/04/2015, Genova, Italy, ISBN: 978-88-98091-38-6. [12] E. Cavalli, A. Lorenzi, “Metodologia e tecnologia per l’e-learning”, Atti del XXXVIII Congresso AICA, pp.759-770, 2000,Taormina, Italy. [13] Convegno, 2 March 2010, E-learning day 2010 ,Centro “Rete Puglia”,Università degli Studi di Bari, Italy. [14] D. Colombo,“Formazione a distanza – ambienti e piattaforme telematiche a confronto”, 2001.

[2] S. Battigelli, A. M. Sugliano, “Archiviazione e condivisione di lesson plan: metadata e applicazioni web 2.0”, 2009, Journal of e-Learning and Knowledge Society, SIEL.

[15] CNIPA, “Il software open source per l'e-learning”, 5 April 2006.

[3] T. Berners-Lee, J. Hendler, O. Lassila, The Semantic Web, 2001, In Scientific American.

[16] S. Dawson, “A study of the relationship between student social networks and sense of community”, Educational Technology & Society, vol. 11, no. 3, pp. 224-238, 2008.

[4] G. Bonaiuti, E-learning 2.0, 2006, Edizioni Erikson. [5] S. Brown, Open and distance learning: case studies from industry and education, 1997, Kogan, London. [6] J. S. Bruner, The Process of Education, University press, 1960, Harvard. [7] P. Campanella, “NetLearn2.0: Piattaforma elearning e metodologie integrative”, Atti Didamatica 2013, Tecnologie e Metodi per la Didattica del Futuro, 07-08-09/05/2013, Pisa, ISBN: 978-8898091-10-2. [8] P. Campanella, “Platforms and methods for the integrated use of educational resources in the processes of e-learning”, ED-MEDIA 2011, World Conference on Educational Multimedia, Hypermedia and Telecommunications 2011, Chesapeake AACE, In T. Bastiaens & M. Ebner (Eds.), Proceedings of World Conference on Educational Multimedia, Hypermedia and Telecommunications 2011, Chesapeake, VA: AACE, 27-28-29-30/0601/07/2011, pp. 2375-2384, Lisbona, Portogallo, ISBN: 18800948904. [9] P. Campanella, “Oracle i-Learning Platform: Un caso di Studio”, Atti DIDAMATICA 2014 – Nuovi Processi e Paradigmi per la Didattica, 07-0809/05/2014, Napoli, Italy, ISBN: 978-88-98091-317. [10] P. Campanella, “Piattaforme per l’Uso Integrato di Risorse Formative nei Processi di e-learning”, Atti DIDAMATICA 2015 – Studio Ergo Lavoro – dalla società della conoscenza alla società delle competenze, 15-16-17/04/2015, Genova, Italy, ISBN: 978-88-98091-38-6. [11] P. Campanella, “Piattaforme proprietarie: Un’analisi metodologica”, Atti DIDAMATICA 2015 – Studio

ISBN: 978-1-941968-26-0 ©2015 SDIWC

[17] E. Demidova, P. Karger, D. Olmedilla, S. Terrier, E. Duval, M. Dicerto, C. Mendez, K. Stefanov, “Services for Knowledge Resource Sharing & Management in an Open Source Infrastructure for Lifelong Competence Development”, 2007, ICALT. [18] V. Di Lecce, A. Giove, M. Calabrese, Ariann@ project: “E-Learning Platform for University Guidance”, In Proceedings of the International Conference on Computer Aided Learning (ICL 08), 24-26 Sept. 2008, Villach, Austria, ISBN: 978-389958-353-3. [19] S. Downes, “E-learning 2.0”, eLearn Magazine, 2005. [20] P. Farace, “Strategie nell’e-learning: l’impatto del modello open source nelle scelte tecnologiche e funzionali”, 2003, http://www.farace.it/files/tesi701482.doc. [21] Fountopoulos, “RichTags: A Social Semantic Tagging System”, Thesis PhD: A dissertation submitted in partial fulfillment of the degree of MSc Web Technology, 2007. [22] S. Fraccavento, “L'e-learning inteso come fenomeno sociale e di mercato (Le differenti piattaforme e tipologie di apprendimento)”, 2003, http://www.studiotaf.it/teoriemodellifad.htm. [23] A. Fuggetta,“Open source software: an evaluation, Journal of Systems and Software, vol. 66, Issue 1, pp. 1-90, 2003. [24] D. F. Garcia, C. Uria, J. C. Granda, F. J. Suarez, F. Gonzalez, “A Functional Evaluation of the Commercial Platforms and Tools for Synchronous Distance e-Learning”, International Journal of

190


Education and Information Technologies, Volume 1, Issue 2, 2007. [25] J. Greer, G. McCalla, J. Collins, V. Kumar, P. Meagher, J. Vassileva, “Supporting Peer Help and Collaboration in Distributed Workplace Environments”, International Journal of Artificial Intelligence in Education, vol. 9, pp. 159177, 1998. [26] S. R. Hiltz, “Supporting Collaborative Learning in Asynchronous Learning Networks”, Open University Symposium on Virtual Learning Environments, 28 April 1997, England. [27] S. Impedovo, P. Campanella, G. Facchini, G. Pirlo, R. Modugno, L. Sarcinella, “Learning Management Systems: un’analisi comparativa delle piattaforme open-source e proprietarie”, DIDAMATICA 2011 Informatica per la didattica, 04-05-06/05/2011, Torino, Italy, ISBN: 9788890540622. [28] S. Impedovo, IAPR Fellow, IEEE S. M., P. Campanella, “DoceboCloud: Apprendimento e Nuove Tecnologie”, Atti DIDAMATICA 2012 Informatica per la didattica, Taranto 14-1516/05/2012, ISBN: 978-88-905406-7-7. [29] M. Lazzari, A. Betella, “Un ambiente open source per la gestione del podcasting e una sua applicazione alla didattica”, In Atti di Didamatica 2007, 10-12 maggio 2007, Cesena, Italy. [30] R. Liscia, “E-learning. Stato dell’arte e prospettive di sviluppo”, 2004, Apogeo. [31] S. Locatelli, “ILIAS Open Source”, In Convegno Elearning e Open Source, 5 dicembre 2003, Roma, Italy. [32] S. Luciani, “Caratteristiche tecniche e funzionalità didattiche delle piattaforme per l’apprendimento online”, 2005. http://www.wbt.it/index.php?risorsa=piattaforme apprendimento [33] R. Mason and F. Rennie, “E-learning and Social Networking Handbook”, 2008, Routledge. [34] V. Mobilio, “Sperimentare piattaforme Open Source: idee per una metodologia d’analisi”, September 2006, In eLearning& Knowledge Management. [35] M. Mobilio, “Moodle, piattaforma Open Source per l’e-learning”, aprile 2008. L’esperienza del CASPUR. [36] R. Palloff, K. Pratt, “Collaborating online: Learning together in community”, 2004, Jossey-Bass. [37] M. Pedroni, “Dall’interoperabilità delle piattaforme all’integrabilità dei moduli interattivi”, Omniacom Editore, Didamatica, pp. 731-735, 2004.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

[38] M. S. Pérez, P. Santos, A. Moghnieh, D. HernándezLeo & J. Blat, October 2008, A Conceptual Framework for the integration of the 2.0 Web Tools in Lifelong Learning Scenarios TENCompetence–Sofia, workshop. [39] T. O’Reilly, “What is the Web 2.0, Design patterns and business models for the next generation of software”, International Journal of Digital Economics, vol. 65, pp. 17-37, 2007. [40] B. Simon, D. Massart, F. Assche, S. Ternier, E. Duval, S. Brantner, D. Olmedilla, Z. Miklós, “A Simple Query Interface for Interoperable Learning Repositories”, In Proc. of the Workshop Interoperability of Web-Based Educational Systems, 2005. [41] L. Stojanovic, S. Staab, R. Studer, “E-learning based on the semantic web”, World Conference on the WWW and Internet, 2001, Orlando, Florida, USA. [42] Z. Yang, Q. Liu, “Research and development of web-based virtual online classroom”, Computer & Education, vol. 48, pp. 171-184, 2008. [43] C. M. Au Yeung, N. Gibbins, N. Shadbolt, “Understanding the Semantics of Ambiguous Tags in Folksonomies”, International Workshop on Emergent Semantics and Ontology Evolution (ESOE2007), 12 November 2007, Busan, South Korea. [44] ADA Lynx (Italia) http://www.lynxlab.com/ada/ada_it.php. [45] Adobe - Information of Connect Pro tool http://www.adobe.com/products/acrobatconnectpro/ [46] aTutor Università di Toronto http://www.atutor.ca/atutor/index.php. [47] Centra - Information of the Centra Symposium tool http://www.centra.com. [48] Claroline Università di Lovanio http://www.claroline.org. [49] Cisco - MeetingPlace Web conferencing solution http://www.cisco.com. [50] Docebo - http://www.docebo.org/. [51] EifFE-L - www.eiffe-l.org/. [52] e/pop - Information of Web Conferencing tool http://www.wiredred.com/web-conferencing/ [53] Freelearn - http://www.freelearn.it/. [54] Elluminate - Information on Elluminate Live tool http://www.elluminate.com. [55] ILIAS - Università di Colonia - http://www.ilias.de. [56] Ilinc - Information of Learn & Conference Linc tools - http://www.ilinc.com. [57] Interwise - Information of the Connect tool http://www.interwise.com. [58] Groove - Information of the Virtual Office tool http://www.groove.net. [59] HotConference - Information on HotConference tool - http://www.hotconference.com. [60] Microsoft - Information of the LiveMeeting tool http://www.livemeeting.com.

191


[61] Marratech - Information on Marratech tool http://www.marratech.com. [62] .LRN - MIT’s Sloan School of Management http://www.dotlrn.org. [63] MegaMeeting - Information of the MegaMeeting tool - http://www.megameeting.com. [64] Moodle - Martin Dougiamas (Australia) http://moodle.org. [65] Raindance - Information on Raindance tools http://www.raindance.com. [66] VoxWire - Information of the Web Conferencing tool - http://www.voxwire.com. [67] WebConference - Features of WebConference tool http://www.webconference.com. [68] WebEx - Information of the WebEx Training Center tool - http://www.webex.com. [69] ACM eLearn - http://www.elearnmag.org/ [70] Journal of Interactive Learning Research (JILR) http://www.aace.org/pubs/jilr/toc.html. [71] Journal of Asynchronous Learning Networks http://www.aln.org/publications/jaln/index.asp.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

192


A Comparative Assessment of e-Learning Platforms P. Campanella Department of Computer Science University of Bari "Aldo Moro" Via Orabona, 4 – 70126 – Bari (Italy) [email protected] ABSTRACT This paper presents a comparative analysis among elearning platforms. Some of the most widespread opensource and proprietary platforms are presented and evaluated. The results, obtained by comparing 28 elearning platforms, highlight the most significant gaps between open-source and commercial platforms, with particular reference to the tools and services supporting social networking and web 2.0 functionalities.

KEYWORDS collaborative learning; e-learning 2.0; distance elearning; e-learning tools; social networks; functional evaluation.

1 INTRODUCTION In the last years, the employment of e-learning technologies is increased considerably, making it possible the development of new approaches for distance learning, based on customizable courseware and activities. In this scenario, the role of e-learning platforms is crucial [31]. The tools and services supported by e-learning platforms determine the possibility to implement alternative learning processes, according to personalized and adaptive strategies [8,21]. In particular, some of the most innovative technologies of web 2.0, define the typical features of the new e-learning 2.0, which specifically concern with the intrinsically social nature of the network [5,13,20]. From the use of simple courseware it has passed to a multitude of services in which e-learners dynamically interact each other using technologically enhanced tools and services, according to the collaborative learning [13,18,31,40]. In this direction, much has been made so far in consideration that most modern

ISBN: 978-1-941968-26-0 ©2015 SDIWC

theories consider the priority and increasingly pervasive aspects of informal learning, at the expense of the old conceptions [3,23,47]. In particular [20], e-learning 2.0 should be able to integrate all the new techniques of learning built around the web 2.0 tools to allow users to build an effective “conversation” on the network [15,33,46]. The technologies of the web 2.0 should allow the development of customized and flexible tools and services, able to support the realization of true virtual learning environments [7,9,23]. With these new services is therefore possible for both teachers and students to share the available contents through blog, podcast, media sharing and social bookmarking [31]. The latter, in particular, with the aim to share bookmarks between different users, even with the ability to add a brief description and keywords with links, so that they are immediately accessible and to create an effective tagging [4,22,43,44,45]. In this scenario it is very difficult to achieve efficient and shared evaluation of elearning platforms. In fact, although numerous attempts have already been described in the literature, difficulty in the evaluation of e-learning platforms derives from the fact that they have undergone a extraordinary evolution in the last years and specifically in concomitance with the development of web 2.0. Furthermore, another problem is that different application domains, in which platforms must be used, have multiple and often conflicting requirements. It is thus quite clear as the analysis in the literature can only partially obtain an objective assessment of the platforms and how these support the learning process, considering the characteristics, needs and issues with particular reference to the use of typical instruments of web

193


2.0 [9,16,34]. This paper presents an analysis of elearning platform addressing the specific functionality related to the use of typical web 2.0 tools and services for collaborative learning [11,14,42]. Particular attention has been focused on the analysis of the two categories: “open-source” and commercial platforms [31,38,39]. Section 2 presents a description of the main features characterizing the two categories of platforms. A detailed analysis of the platforms, which takes in consideration 28 of the most popular e-learning platforms (10 open-source and 18 commercial), is given in section 3. A short discussion of the results obtained and the conclusion is reported in the section 4. 2 PLATFORMS E-LEARNING: A COMPARATIVE EVALUATION This section provides a comparative evaluation of the main platforms for e-learning, with particular reference to their capacity to support services well known and widespread, typical of web 2.0, and then the “social network” such as facebook, feedback, chat, blog, forum, youtube, podcasting, skype and wiki [5,10,19,20,24,32,37]. The tools are: facebook: it is a social utility that connects people with friends and others who work, study and live around them. youtube: tools share common among users of social web in interpersonal as well as community. skype: freeware proprietary tools for instant messaging and VOIP, even in a peer to peer. The services are: feedback: service used in inter-communication and sharing of knowledge (questionnaires, interviews, focus groups). chat: synchronous service used to promote individual and group communication in the form of intra-community. blog: tool intercommunication online learning, content sharing within the community.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

instrument used in virtual forum: intercommunity chat is that unlike the asynchronous. podcasting: direct service to the production, network sharing and use of audio and/or video through the use of multiple technologies, partly online and partly online wiki: service that allows you to share collaboratively creating new knowledge. Figure 1 and 2 show in particular the analysis of ten between the main open-source platforms: ada, atutor, claroline, docebo, eiffe-L, freelearn, ilias, moodle2, .LRN, sakai. Is evident that some services such as feedback, blog, forum, podcasting and wiki are supported by all platforms, whereas no opensource platform supports skype and youtube (Fig. 1). Concerning facebook, it is only supported by the platforms docebo, moodle2 and .LRN, whereas the platform .LRN is the only one not supporting the chat (Fig. 2).

Figure 1. Open-Source Platforms: Tools

Figure 2. Open-Source Platforms: Services

Fig. 3 and 4 show the analysis of eighteen proprietary platforms, including the most popular

194


and well known: adobe connect, centra, elluminate live, e/pop, groove, hotconference, megameeting, netlearning, oracle ilearning, picturetalk, raindance, saba learning enterprise, same time, t-learn, voxware, webct, wave three, webconferencing. They show in particular that all proprietary platforms considered support the services such as feedback, chat, blog, forum, podcasting and wiki (Figure 4), but none of them supports the tools facebook, youtube and skype (Fig. 3).

Figure 3. Proprietary Platforms: Tools

platforms is visible an effort toward the integration of tools for the development of social networks [11,35,36,40]. The results also clarify that most of the platforms allows the use of an off-line tool for creating courses, and modify and update contents. 3 SUMMARY EVALUATION PLATFORMS The results of the previous section explain some aspects of the current phenomenon of migration from static to dynamic learning approaches. In other words, the advancements of e-learning platforms show that static approaches, which are in general considered by commercial platforms, are now evolving towards personal learning spaces, in which the learning procedures are not limited to the duration of a course, but they accompany the student also after the formal and circumscribed phase of learning, according to the lifelong learning principles [6,11,18,42]. This phenomenon also suggests the reasons for which multimedia content delivery, high storage requirements and computational power, along with strong constraints in the budget, frequently leads companies and institutions to find e-learning solutions in outsourcing [6,12,37]. In order to analyze the critical factor of success of e-learning initiatives, researchers focus on supporting the collaborative factors that are identified as follows [14,28,31]: ° use of Open Source (OS) technologies; ° use peer-to-Peer (P2P) strategies for finding contents; ° use Virtual Learning Community (VLC) for learning.

Figure 4. Proprietary Platforms: Services

The results show that, in less than a few products, there is no substantial difference between the platforms, as the market tends to make them comparable. An important consideration that must be stressed is that widely distributed tools, like youtube and skype, are not integrated into any of the tested platforms, neither open-source nor commercial [15,17,41]. Anyway, in open source

ISBN: 978-1-941968-26-0 ©2015 SDIWC

OS technologies are considered as “the world that creates value” in a collaborative manner. Thanks to free access to source code, individuals work together to customize the online learning systems. Similarly, P2P systems, prime example of collaboration between computers, allow you to easily expand the network of participants by offering greater availability of resources. Finally, the VLC, the new models that represent intra and

195


inter-organizational support online learning, enabling individual participants to be fully interactive [2,26,29,30]. These three elements, with their mutual interrelations, generate the paradigm of “Learning Collaborative” that combines aspects of learning, organizational and technological learning process [12,25,27]. In light of this latest research platforms can be divided into three different categories: [1] platforms designed mainly to provide content and learning units (courses, classes, module) by enjoy exclusively over the Internet. This category includes almost all the trading platforms. This type of platform have the ability to support very high loads of users; [2] platforms whose primary function is to be a virtual bridge between the teacher and the student. In these systems, which represent the majority of open-source platforms, usability is of paramount importance in order to encourage the proper use of diverse resources according to the specific user needs (such as modules for chat, video delivery, questionnaires and collaborative activities). The use of these platforms is generally in mixed mode: on-line and off-line. The great advantages of these platforms is the possibility of being customizable and free of charge; [3] platforms for collaborative learning, in which the difference between student and teacher tends to disappear and that is even more emphasized the use of forums and chat. 4 DISCUSSION AND CONCLUSIONS Following the increasing demand from new categories of users, the availability of e-learning platforms is continuously increasing in terms of both commercial systems and open-source solutions. In this scenario, platform evaluation is often difficult since their assessment strongly depends on the specific requirements of the application domain. Notwithstanding, platform evaluation is very important since they often

ISBN: 978-1-941968-26-0 ©2015 SDIWC

represent a factor of paramount importance for the success of the e-learning activities. This paper has the aim to provide a useful contribution towards platform assessment. In particular the evaluation concerns specific aspects related to the development 2.0 tools and services. The results show that, given the large number of solutions available, only very few platforms support typical instruments of social networks. In addition, some cases are exists in which platforms are flexible enough to set the content in a way suitable for the specific educational needs, for example through the use of additional modules or plug-in. In most cases however, an analysis of the peculiarities of the platforms, both proprietary and open-source, shows that the lack of flexibility is a characteristic common to many platforms. This is particularly true in the case of proprietary platforms. In this case although several products on the market exists, their difference is actually rather limited since all the solutions are very similar. In conclusion, it can be argued that the analysis highlights the rigidity of elearning platforms currently available. In general, platforms are still far from having the characteristics of web 2.0 and social networking. This is specifically true for proprietary platforms and for many open-source solutions. Conversely, there are only some open source platforms that can show advanced characteristics, since they are supported by a large community that update and enhance their capabilities in the direction of the requirements of web 2.0. Therefore, the progress of research in this area is still necessary, in order to improve state of art platforms to meet emerging learning requirements, such as the collaborative creation and editing of educational resources through advanced social tagging strategies and unstructured educational resources exploitation. REFERENCES [1] ACM eLearn - http://www.elearnmag.org/ [2] R. C. Atkinson, R. M. Shiffrin, “A proposed System and its Control Process”, In Psychological of Learning and

196


Motivation: Advances in Research and Theory, vol. 2, 1968, pp. 89-105, New York Academic press. [3] S. Battigelli, A. M Sugliano, “Archiviazione e condivisione di lesson plan: metadata e applicazioni web 2.0”, Journal of e-Learning and Knowledge Society, SIEL, 2009. [4] T. Berners-Lee, J. Hendler, O. Lassila, “The Semantic Web”, In Scientific American, 2001.

[15] D. Colombo, “Formazione a distanza – ambienti e piattaforme telematiche a confronto”, 2001, http://www.irre.lombardia.it/TD/FAD/ricerca_livello3p.ht m. [16] CNIPA, “Il software open source per l'e-learning”, 5 April 2006. [17] S. Dawson, “A study of the relationship between student social networks and sense of community”, Educational Technology & Society, Vol. 11, No. 3, 2008, pp. 224-238.

[5] G. Bonaiuti, “E-learning 2.0”, Edizioni Erikson, 2006. [6] S. Brown, “Open and distance learning: case studies from industry and education”, Kogan, London, 1997. [7] J. S. Bruner, “The Process of Education”, University press, Harvard, 1960. [8] P. Campanella, “NetLearn2.0: Piattaforma e-learning e metodologie integrative”, Atti Didamatica 2013, Tecnologie e Metodi per la Didattica del Futuro, 07-0809/05/2013, Pisa, ISBN: 978-88-98091-10-2. [9] P. Campanella, “Platforms and methods for the integrated use of educational resources in the processes of elearning”, ED-MEDIA 2011, World Conference on Educational Multimedia, Hypermedia and Telecommunications 2011, Chesapeake AACE, In T. Bastiaens & M. Ebner (Eds.), Proceedings of World Conference on Educational Multimedia, Hypermedia and Telecommunications 2011, Chesapeake, VA: AACE, 2728-29-30/06-01/07/2011, pp. 2375-2384, Lisbona, Portogallo, ISBN: 18800948904. [10] P. Campanella, “Oracle i-Learning Platform: Un caso di Studio”, Atti DIDAMATICA 2014 – Nuovi Processi e Paradigmi per la Didattica, 07-08-09/05/2014, Napoli, Italy, ISBN: 978-88-98091-31-7. [11] P. Campanella, “Piattaforme per l’Uso Integrato di Risorse Formative nei Processi di e-learning”, Atti DIDAMATICA 2015 – Studio Ergo Lavoro – dalla società della conoscenza alla società delle competenze, 15-16-17/04/2015, Genova, Italy, ISBN: 978-88-9809138-6. [12] P. Campanella, “Piattaforme proprietarie: Un’analisi metodologica”, Atti DIDAMATICA 2015 – Studio Ergo Lavoro – dalla società della conoscenza alla società delle competenze, 15-16-17/04/2015, Genova, Italy, ISBN: 978-88-98091-38-6. [13] E. Cavalli, A. Lorenzi, “Metodologia e tecnologia per l’elearning”, Atti del XXXVIII Congresso AICA, 2000, pp.759-770, Taormina, Italy. [14] Convegno, 2 March 2010, E-learning day 2010 ,Centro “Rete Puglia”,Università degli Studi di Bari “Aldo Moro”, Italy.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

[18] E. Demidova, P. Karger, D. Olmedilla, S. Ternier, E. Duval, M. Dicerto, C. Mendez, K. Stefanov, “Services for Knowledge Resource Sharing & Management in an Open Source Infrastructure for Lifelong Competence Development”, 2007, ICALT. [19] V. Di Lecce, A. Giove, M. Calabrese, “E-Learning Platform for University Guidance”, In Proceedings of the International Conference on Computer Aided Learning (ICL 08), Villach, Austria, 2008. [20] S. Downes, “E-learning 2.0”, eLearn Magazine 2005. [21] P. Farace, “Strategie nell’e-learning: l’impatto del modello open source nelle scelte tecnologiche e funzionali”,2003, http://www.farace.it/files/tesi701482.doc. [22] Fountopoulos “RichTags: A Social Semantic Tagging System”. Thesis PhD: A dissertation submitted in partial fulfillment of the degree of MSc Web Technology, 2007. [23] S. Fraccavento,“L'e-learning inteso come fenomeno sociale e di mercato (Le differenti piattaforme e tipologie di apprendimento)”, 2003, http://www.studiotaf.it/teoriemodellifad.htm. [24] Fuggetta, “Open source software: an evaluation”, Journal of Systems and Software, Volume 66, Issue 1, 2003, pages 1-90. [25] D. F. Garcia, C. Uria, J. C. Granda, F. J. Suarez, F. Gonzalez, “A Functional Evaluation of the Commercial Platforms and Tools for Synchronous Distance e Learning”, International Journal of Education and Information Technologies, Volume 1, Issue 2, 2007. [26] J. Greer, G. McCalla, J. Collins, V. Kumar, P. Meagher, J. Vassileva, “Supporting Peer Help and Collaboration in Distributed Workplace Environments”, International Journal of Artificial Intelligence in Education, vol. 9, 1998, pp. 159-177. [27] S. R. Hiltz, “Supporting Collaborative Learning in Asynchronous Learning Networks”, 28 April 1997, Open University Symposium on Virtual Learning Environments, England.

197


[28] IRRE Lombardia, “E-learning Integrato su Open Source, percorso di ricerca e sperimentazione di piattaforme Open Source per l'e-learning”, 2004, Progetto E.L.I.O.S., http://www.irre.lombardia.it/TD/FAD/tecnologie.htm. [29] Journal of Interactive Learning Research (JILR) http://www.aace.org/pubs/jilr/toc.html. [30] Journal of Asynchronous Learning Networks http://www.aln.org/publications/jaln/index.asp.

-

[31] S. Impedovo, P. Campanella, G. Facchini, G. Pirlo, R. Modugno, L. Sarcinella, “Learning Management Systems: un’analisi comparativa delle piattaforme open-source e proprietarie”, DIDAMATICA 2011 - Informatica per la didattica, 04-05-06/05/2011, Torino, Italy, ISBN: 9788890540622.

the integration of the 2.0 Web Tools in Lifelong Learning Scenarios TENCompetence–Sofia”, workshop. [43] T. O’Reilly, “What is the Web 2.0”. Design patterns and business models for the next generation of software, International Journal of Digital Economics, Vol. 65, pp. 17-37, 2007. [44] B. Simon, D. Massart, F. Assche, S. Ternier, E. Duval, S. Brantner, D. Olmedilla, Z. Miklós, “A Simple Query Interface for Interoperable Learning Repositories”. In Proc. of the Workshop Interoperability of Web-Based Educational Systems. [45] L. Stojanovic, S. Staab, R. Studer, “E-learning based on the semantic web”, 2001, World Conference on the WWW and Internet, Orlando, Florida, USA, 2005.

[32] S. Impedovo, IAPR Fellow, IEEE S. M., P. Campanella, “DoceboCloud: Apprendimento e Nuove Tecnologie”, Atti DIDAMATICA 2012 - Informatica per la didattica, Taranto 14-15-16/05/2012, ISBN: 978-88-905406-7-7.

[46] Z. Yang, Q. Liu, “Research and development of webbased virtual online classroom”, Computer&Education, vol. 48, pp. 171-184, 2008.

[33] M. Lazzari, A. Betella, “Un ambiente open source per la gestione del podcasting e una sua applicazione alla didattica”, 10-12 maggio 2007, In Atti di Didamatica 2007, Cesena, Italy.

[47] C. M. Au Yeung, N. Gibbins, N. Shadbolt, “Understanding the Semantics of Ambiguous”, 12 November 2007, Tags in Folksonomies, International Workshop on Emergent Semantics and Ontology Evolution (ESOE2007), Busan, South Korea.

[34] R. Liscia, “E-learning. Stato dell’arte e prospettive di sviluppo”, 2004, Apogeo. [35] S. Locatelli, “ILIAS Open Source”, 5 dicembre 2003, In Convegno E-learning e Open Source, Roma, Italy. [36] S. Luciani, “Caratteristiche tecniche e funzionalità didattiche delle piattaforme per l’apprendimento online”,2005, http://www.wbt.it/index.php?risorsa=piattaforme apprendimento. [37] R. Mason and F. Rennie, “E-learning and Social Networking Handbook”, 2008, Routledge. [38] V. Mobilio, “Sperimentare piattaforme Open Source: idee per una metodologia d’analisi”, September 2006, In eLearning & Knowledge Management. [39] M. Mobilio, “Moodle, piattaforma Open Source per l’elearning”, aprile 2008, L’esperienza del CASPUR. [40] R. Palloff, K.. Pratt, “Collaborating online: Learning together in community”, 2004, Jossey-Bass. [41] M. Pedroni, “Dall’interoperabilità delle piattaforme all’integrabilità dei moduli interattivi”, Omniacom Editore, Didamatica, 2004, pp. 731-735. [42] M. S. Pérez, P. Santos, A. Moghnieh, Hernández- D. Leo & J. Blat, October 2008, “A Conceptual Framework for

ISBN: 978-1-941968-26-0 ©2015 SDIWC

198


EDA as a Discriminate Feature in Computation of Mental Stress Khalid Masood Faculty of Computing and IT, University of Jeddah, Saudi Arabia. E-mail: [email protected] Abstract—In computation of mental stress, various features are determined from a range of physiological signals. During stress, hormones levels inside the body of a stressed person are changed that results in a number of biomedical signals that are communicated among different body organs. A wireless wearable platform has been designed that record these biomedical signals. To induce stress, a series of cognitive experiments were developed that produce stress on the participants. EDA, HRV, respiration and brain signals are used for computing features and the objective was to identify most significant feature or their various combinations. It is verified that EDA features achieves a similar accuracy that can be obtained using various combination of features or using a master set containing all the features. The classification accuracy is more than 80% using EDA with a SVM model containing rbf kernel. Keywords-E-health; mental stress; signals; EDA; SVM; wireless sensors

bionedical

INTRODUCTION In response to a dangerous situations or a threat, the brain of a human body makes necessary arrangements to cope with the challenge. The sense of that unease form the normal physical conditions is defined as stress [1]. In unexpected situations containing challenges, the nervous system of a person is activated and hormones are released to counter affect that threat. Stress is experienced as when the demand of external or environmental factors exceeds a person’s ability to cope with and control these factors [2]. Stress is also characterized by environmental conditions in which people face high demands, but have little control or influence over their external environments [3]. In the studies for stress, it is reported that short periods of stress results in reactions that are damaging for a healthy life such as disturbance in sleep, changes in mood, headache and stomach disorder etc. When stress is prolonged, a wide range of mental and physical health problems emerge including depression, anxiety, cardiovascular diseases, high blood pressure and thoughts of suicidal attempts etc. Other factors that stress can I.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

contribute in a daily life are emotional strain and reduction in the quality of life by affected persons. Also there is a significant financial burden on an individual who is coping with stress as it would cost him a reasonable amount to pay for the bills of medical and insurance cover [4]. The Autonomic Nervous System (ANS) which is responsible for response to a stressor, is composed of the Sympathetic Nervous System (SNS) and the Parasympathetic Nervous System (PNS) [5]. Activities in SNS increase during stress conditions and in resting periods, PNS dominates and brings body conditions back to normal. SNS and PNS control the physiological measures such as heart rate variability (HRV), electrodermal activity (EDA) and brain activity (EEG) which are primary signal for computing stress [6]. There are other physiological activities that also activate ANS and necessary precautions should be adapted to separate physical activities’ impact from stress related activities [7]. The hormones system becomes active in response to psychological stress. In chronic stress, two types of hormones, HPA and SAM systems are triggered repeatedly and remain active on prolonged basis. Thus they interfere with the release and control of other physiological system which in turn increases the risks of psychological and physical disorders. In this study, we have identified the most discriminate feature during computation of mental stress. A model to compute stress has been developed that contains wearable wireless sensors to record various physiological parameters. These parameters are changed when a stressor is applied onto a human body. In controlled laboratory conditions, various cognitive tests are performed which act like real life stressors for inducing stress in the participants. In response to stressor, physiological signals that contain HRV, EDA, respiratory affects and brain signals are changed that are recorded. During and in between mental activities, relaxing conditions are applied to

199


relieve the effects of stress. A number of cognitive activities have been designed to produce different levels of stress. A feature selection technique is tested on the obtained features to compute the discriminatory power of features. Classification is performed on individual features and their various combinations. Based on the accuracy, the most discriminate feature is identified along with its various combinations. HARDWARE Heart rate activity can be recorded by various devices. The gold standard for recording is ECG but it requires two electrodes and wiring which is unsuitable for long term measurements [8]. Pulse oximetry can also be used to measure HRM but it is very sensitive to motion artifacts. The optimal solution is heart rate monitor (HRM) that is used to record heart rate variations in cardiovascular activities or in stressor’s response [9]. It contains a strap which is worn around the chest. A wireless transmitter is connected to the strap that transmits heart rate to holster unit. Polar Elecro Inc. manufactures Polar Wearlink HRM which was used in our experiments. In Figure 1, a human body is presented with the complete wearable sensor platform. II.

Respiration contributes significantly in heart rate variations and there is a need to record respiration effects. To monitor breathing effects, a variety of sensing technologies can be used. The variation in abdominal cross section or thoracic is measured by Respiratory Inductive Pletthysmography (RIP) using an abdominal strap that measures changes in the magnetic fields of embedded coils [10]. Placing two electrodes in the rib cage that records impedance changes in the alternating current variations due to respiration are recorded in Impedance Pneumography (IP). For long term monitoring, both these sensors are unsuitable due to postural changes and motion artifacts. In our study, a pressure based respiration sensor manufactured by Thought Technology Ltd. (SA9311M). It is insensitive to motion artifacts and is easily integrated to chest strap with HRM. In Figure 2 a holster unit is presented that contains data process unit with a sensor hub and a lithium ploymer battery that can provide continuous power for about 13 hours.

Figure 2: A holster unit.

Figure 1: A human wearing a chest strap, an abdomen strap and AgCL electrodes in the fingers.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

An electrical voltage of low level is applied to the skin to monitor changes in skin conductance using electrodermal activity (EDA) [11]. In case of stress, body glands release sweat in palms and fingers which in turn increases the skin conductance. EDA can be monitored in palms of the hands but for long term use, they are unsuitable. In our experiments, two AgCl electrodes are attached to middle and index fingers of non-dominant hand to record skin conductance. These electrodes are made by Vivo Metric Systems Corp. (E243). In Figure 2,

200


a PC USB transducer is shown to record sensor data in real time on a PC server. An abdominal strap contains holster unit with three components integrated to it. The components include a data processing unit, a sensor hub and a battery. A 2 GB mini SD flash card is used for data storage and is mounted on a Vertex Pro motherboard with 400 MHz processing speed (Gunstix Inc.). A sensor hub is also connected to the holster unit which is made up of a 3D accelerometer from STMicroelectronics, a GPS unit form Linx Technologies Inc. along-with a clock unit from Dallas semiconductor Inc. A HRM receiver module is also connected to sensor hub alongwith a wireless transceiver used for communication with wireless sensors. A built in charging module is attached in the sensor hub that charges the 3000 mAh Li-Po battery that can be used for continuous data collection up to thirteen hours. It contains HRM in the chest strap to monitor heart rate variation, two electrodes in the fingers for EDA sensor to monitor skin conductance and holster unit in the abdomen for transmitting and storing data [12]. COGNITIVE EXPERIMENTS There are 24 participants containing equal number of male and female subjects. A medical doctor examined the physical health of the participants and each participant provided his/her written consent on the forms. The experimental procedure was briefed to each subject and he/she was not trained for any of the mental activities. . In Figure 3, a sequence is shown for the experimental protocol. It starts with deep breathing to initialize with normal conditions. A mental challenge is followed that induces a pre determined level of stress based on the severity and difficulty of the task. At the end of each activity, deep breathing exercise is performed repeatedly to relieve the body from the effects of stress. III.

To assist the experiments, a protocol was designed to induce mental stress in controlled indoor conditions. There are six deep breathing exercises and five mental challenges for the participants. First of all, the system is calibrated for each individual and an initial deep breathing activity is performed to form a baseline. Each deep breathing session is performed for three minutes. In that session, a subject has to take breathes or inhale for 4 seconds and then breathe out or exhale for 6 seconds. The procedure is repeated and continued for 3 minutes. After the first deep breathing session, a mental challenge of memory search has to be performed by the subject. There is another deep breathing session after each mental challenge to relax the subject and bring back the body to a normal condition. The second mental challenge was color word test that lasted for 5 minutes. A 3rd deep breathing session was performed again to prepare the subject for the next challenge. Next challenges consist of mirror trace, dual task and public speech. The duration of each challenge was 5 minutes. At the end, a final session of deep breathing is performed. Subjects had to rate each mental challenge with various difficulty levels following a Linkert scale, where a minimum difficulty is rated as 1 and extreme stressful challenge is rated as 7. In Figure 4, a screenshot for color word test (CWT) is presented. The user has to respond on sound, text or bar color to progress in the challenge. There are random questions to determine the color based on sound, typing or shown color of a bar. The user gets confused as there is a very little time to concentrate what form of question he/she has to reply as sound, picture and word, all three are depicting different colors.

Figure 3: A sequence of activities that starts with normal conditions followed by a mental challenge to induce stress and finally bringing back body to the normal conditions. Figure 4: A screenshot for color word test (CWT).

ISBN: 978-1-941968-26-0 ©2015 SDIWC

201


EDA FEATURE SELECTION Six parameters are extracted from physiological signals obtained by various wireless sensors. At 500 Mhz, heart rate signals were sampled using a peak detection algorithm. The resulting signal was re-sampled at 4 Hz. Very low frequency (VLF) component was removed from the signal using a band pass filter between 0.04 Hz and 0.4 Hz. Four features were extracted from heart rate variation (HRV) analysis [13]. First extracted feature was AVNN which was an average of time interval between normal heart beats. The second feature was pNN25 which showed the percentage difference greater than 25 msec for adjacent NN intervals. The 3rd feature was root mean square of successive difference (RMSDD) and the 4th was HRV-HF for high frequency power of HRV. For respiration, Resp-LF showed low frequency respiratory power. Finally skin conductance was monitored in SCR that recorded few seconds of short time intervals whereas SCL was ignored that captures the skin conductance impedance for longer time periods. The EDA features are selected as representative physiological parameters for the proposed model as they are linearly proportional to stress levels in comparison to HRV features which vary inversely. To form a representative signature for EDA, principal component analysis is performed on EDA features and its first principal component is extracted which contains more than 90% variance of these features. There are two components of EDA. Skin conductance level (SCL) is the slowly changing offset and skin conductance response (SCR) is a series of transient peaks. Two features, mean and standard deviation, are computed from SCL as follows, IV.

where is the average SCL trend for N samples of signature . The standard deviation is computed as follows, ]1/2 where is standard deviation of the conductance signature . Similarly and are computed from residual SCR. In our experiments, SCL is used as it maps accurately

ISBN: 978-1-941968-26-0 ©2015 SDIWC

the perspiration level in a human palm and fingers. There is a relation between brain activity and stress. The functional magnetic resonance imaging (fMRI) and electroencephalography (EEG) are used frequently to analyze brain activity signals while positron emission tomography (PET) is also used but is less popular. Generally EEG is used most commonly as it has high temporal resolution and low cost. A brain generates electrical signals due to neural activities. Using electrodes at scalp, brain signals are recorded in EEG that represents electrical waveforms of complex nature. Electrodes are charged at 20 to 100 micro volts and are placed on both sides of the scalp that contains right and left brain hemispheres. Frequency, amplitude and shape of the scalp are used to identify and analyze waveforms. During negative emotions, activities in the right hemisphere of the brain dominate than the activities that are produced in the left side of the brain. The right hemisphere is thus the area to be explored for determining stress [14]. Frequency and amplitude is used to categorize EEG signals and determine a particular state of a person. Conscious states are presented by beta and alfa waveforms whereas unconscious states are denoted by theta and delta waves. Stress is mainly indicated by increase and rapid growth of beta wave frequencies and at the same time alpha frequencies are decreased. In right handed persons, the amplitude of alpha waves is slightly higher on the non dominant side. The band pass filtering technique is commonly used for analysis of brain signals [15]. Fourier transform and wavelet packets are used to analyze EEG with respect to frequency, time and spatial domains. Stress is computed using the ratio of power spectral densities of alpha waves frequencies along-with beta wave frequencies. Ratios are defined as following,

where αR and αL are alpha bands on the right and left hemispheres of the brain and similarly βR and βL are beta bands. To determine relaxation levels, summation of alpha along-

202


with theta and summation of alpha, beta and theta measures are used. EXPERIMENTAL RESULTS In these experiments, wireless sensors were used to read biomedical signals. There are four sensors that monitor HRV, EDA, respirations and brain signals. There were two classes of activities, mental challenges that act like stressors and induce stress. Other kinds of activities are deep breathing exercises that relax the body. The purpose of experimentation is to compute stress levels that were induced by various mental exercises. A SVM classification model was designed that classify the induced parameters into two classes. One was stress class and other was relax class. From HRV, six features were extracted and EDA was employed for three features. Two features were extracted from respiratory signal and three features were computed from brain signals. Cross validation experiments were designed with four fold cross validation. There were three sets of training data and one set was used for testing. Each time on the next turn, the test set was replaced by a training set and the process keeps on repeating. After tuning the parameters for kernel bandwidth and cost function, individual parameters and their various combinations are employed in an empirical manner. First of all, HRV features are put in the training set and classification accuracy was computed for the test set. Secondly EDA features were put in the experiments. Similarly respiratory and brain signal features were used to find the classification accuracy. In the second phase, all features were combined and classification accuracy is determined. In the next stage, various combinations of the features such as HRV and EDA, EDA and respiratory, EDA, HRV and respiratory etc. are used. The results are presented in Table 1 and Figure 5. V.

CONCLUSIONS AND FUTURE WORK EDA parameters are shown to be most discriminate among all the physiological parameters. The combined feature vector containing all the features from all recorded signals obtained the similar accuracy as was VI.

ISBN: 978-1-941968-26-0 ©2015 SDIWC

achieved by EDA features alone. Although HRV and respiratory features are important also but EDA features are most significant and discriminate as sweating is directly proportional to various stress levels. Table 1: Accuracy chart showing the classification rates for various features and their combinations.

Correct Rate Error Rate Sensitivit y Specificit y

HRV

Respi

EDA

HRV+ EDA 81

All

82

HRV+ R 74

72

70

28

30

18

26

19

17.5

74

72

78

73

78

79

78

76

81

78

82

84

100 80

Correct

60

Error

40

Sensitivity

20

Specificty

0 HRV

Resp

EDA

All

Figure 5: Classification Accuracies chart.

A stress prediction system has been designed that contains wearable wireless sensors that record physiological parameters that vary in response to different stress levels. HRV provides heart rate variations and in stress, heart rate decreases. EDA is the conductance of electrical signals in the fingers of a person and in stress, EDA increases. Similarly respiratory features and brain activity signals also vary stress. To induce stress, a protocol has been designed that contain various mental challenges and different levels of stress are induced when engaged with these activities. Deep breathing is used in between and start of each mental exercise to bring the body back to its normal conditions and relieve the effects of stress. Wireless sensors record the variation in physiological signals and different features are extracted from these signals. For classification,

203

82.5


SVM model is used to differentiate stress conditions from the relax situations. Various combinations of features are employed and it is concluded that EDA features alone achieves almost similar accuracy as divided by using all combined features. In future, the model would be tested on real time stress conditions such as fire fighting scenes and students in the exam centers. In hardware, physiological signals can be recorded with more improved devices. Also, the classification model would be improved by incorporating Bayesian model into the system. ACKNOWLEDGEMENT This work was supported by the Deanship of Scientific Research (DSR), university of Jeddah and King Abdulaziz university, KSA. The author is indebted to Dr Riccardo GuiterrezOsuna and Dr Beena Ahmed as they assisted in providing the data and provided useful comments in preparation of the manuscript. REFERENCES [1] H. Seyle, The Stress of Life, Mcgraw-Hill, 1956. [2] J. P. Niskanen, M. P. Tarvainen, P. O. Ranta-Aho and P. A. Karialainen, "Software for Advanced HRV analysis," Computer Methods and Programs in Biomedicine, vol. 7, no. 6, pp. 73-82, 2004. [3] T. Steckler, Handbook of Stress and the Brain, Amsterdam: Elsevier Science, 2005. [4] J. Zhai and A. Baretto, "Stress recognition using noninvasive technology," Proceedings of 19th International Florida Artificial Intelligence Research Society Conference FLAIRS, pp. 395-400, 2006. [5] L. K. McCorry, "Physiology of the Autonomic Nervous System," American Journal of Pharmaceutical Education, vol. 71, no. 4, pp. 78-85, 2007. [6] S. Bakewell, "The Autonomic Nervous System," World Federation of Socities of Anaesthesiologists, vol. 1, no. 5, pp. 1-2, 1995. [7] J. F. Thayer, S. S. Yamanoto and J. F. Brosschot, "The relaionship of Autonomic Imbalance, Heart rate Variability and Cardiovascular Disease Risk

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Factors," International Journal of Cardiology, vol. 141, no. 2, pp. 122-131, 2010. [8] J. Choi, B. Ahmed and R. Guiterrez-Osuna, "Development and Evaluation of an Ambulatory Stress Monitor Based on Wearable Sensors," IEEE Transactions on Information Technology in Biomedicine, vol. 16, no. 2, pp. 279-286, 2012. [9] D. Bansal, M. Khan and A. K. Salhan, "A Review of Measurement and Analysis of Heart Rate Variability," International Conference on Computer and Automation Engineering (ICCAE'09), pp. 243246, 2009. [10] D. L. Elghazi, D. Laude and A. Girard, "Effects of Respiration on Blood Pressure and Heart rate variability in Humans," Clinical and Experimental Pharmacology and Physiology, vol. 18, no. 11, pp. 735-742, 1991. [11] S. C. Jacobs, R. Friedman, J. D. Parker, G. H. Toffler, A. H. Jimenez, J. E. Muller, H. Benson and P. H. Stone, "Use of Skin conductance changes during mental stress testing on an index of autonomic arousal in cardiovascular research," Ameican Heart Journal, vol. 128, pp. 1170-1177, 1994. [12] U. R. Acharya, K. P. Joseph, N. Kannathal, C. M. Lim and J. S. Suri, "Heart Rate Variability: A Review," Medical and Biological Engineering and Computing, vol. 44, no. 12, pp. 1031-1051, 2006. [13] M. Kumar, M. Weippert, R. Vibrandt, S. Kreuzfeld and R. Stoll, "Fuzzy Evaluation of Heart Rate Signals for Mental Stress Assessment," IEEE Transactions on Fuzzy Systems, vol. 15, no. 5, pp. 791-808, 2007. [14] N. Sulaiman, N. H. Hamid, Z. H. Murat and M. N. Taib, "Initial Investigation of Human Physical Stress Level using Brain Waves," IEEE Student Conference on Research and Development (SCOReD), pp. 230233, 2009. [15] J. L. Burns, E. Labbe, B. Arke, K. Capeless, B. Cooksey, A. Steadman and C. Gonzales, "The Effects of different types of Music on perceived and physiological measures of stress," Journal of Music Therapy, vol. 28, pp. 104-116, 2002.

.

204


A New Approach to Simulate Operation Rooms based on Medical Devices used in Surgical Procedures Sana Ghobadi, Arash Taki, PhD, Mohammad Aziz Esmaeili Islamic Azad University, IAU, UAE Branch [email protected], [email protected] , [email protected] Abstract Patient safety is one of the greatest challenges in healthcare. In the operating rooms (ORs), errors are frequent and often consequential. Medical devices and specially advanced technologies have a key role in improving patient outcomes in ORs. In this paper, a new approach is presented to simulate operating theater based on different types of surgeries and medical technologies used in different surgical cases. In the first step, clinical background of different surgical procedures such as: General, Cardiac, Neuro and Laparoscopic surgeries are collected and analyzed. Then all relevant medical devices of each procedure are categorized based on complexity of technologies and surgical procedures. All technical information such as: physics and technical basics, technical specifications, block diagrams, type of devices, top manufacturers and troubleshooting are gathered. Finally, a graphical user interface (GUI) is developed by C++ to implement all collected data in an interactive application. Keywords: Graphical User Interface (GUI), Surgical Procedure, Operating Room, Biomedical Engineering Introduction The Heart of Hospital is another name given to the Operation Room in Hospitals nowadays. This is because of the number of operations done per annum worldwide. According to a research done by WHO

ISBN: 978-1-941968-26-0 ©2015 SDIWC

(World Health Organization) organization, it was estimated that 234 million major surgical procedures are undertaken every year worldwide which quiet large number [1] [2]. During these surgeries due to some mistakes such as machine errors, surgeon mistakes, or OT technicians mistakes many patients die. According to a research done by IOM (Institute Of Medicine) 32 Thousand patients die per annum worldwide [3] during the surgeries caused by man made errors which is a noticeable number. By considering these matters and studying the reasons for all these failures we have come up with an idea of creating a user friendly program in which the OT technicians, surgeons, and biomedical engineers can learn and trouble shoot the machines within no time. We have also gathered clinical related information as well as technical related information which can be used by technicians, surgeons, and biomedical engineers at the same time. The program which we have designed will make sure the patient safety and ensure the success of operation which is believed that it’s a great innovation in operation rooms. In the year 2010 the first similar application was created by Allis Technology Company called ScrubUp. And they were the first practical mobile resource tool targeted to assist surgical technicians, instrument & circulating nurses. ScrubUp provides preloaded information that aims to guide the surgical nurse/technician in being able to confidently prepare & set up an operating room capturing all the

205


intraoperative details required. Surgical preferences can be updated, changed & reviewed, anywhere, anytime. The ScrubUp application is a one dimensional program whereas our program is a multidimensional.

professors and biomedical students, and their opinions were collected as feedback.

This simulator is an effective tool to train and update all healthcare workers and especially biomedical engineers to utilize medical device technologies in a very efficient and safe way during surgical procedures. This GUI can also be used for educational purpose for medical and biomedical students. This project is the mixture of technical and clinical parts of an operation room. This fills the empty space in an application similar to ScrubUp. We focus on both the technical aspect and clinical aspect in which the technicians, surgeons, and biomedical engineers utilize it. The rest of this paper is structured as follows. Section 2 describes the framework of proposed method for data collection, classification and simulation. Section 3 shows experimental results of the proposed method. Finally Section 4 discusses the advantages of proposed simulator and concludes the paper.

Fig.1-Three steps in this project Step1 In this article, we surveyed 3 operation rooms based on 3 different surgical procedures. Since the majority of surgical operations are general surgeries [4], neurosurgery [5], and cardiovascular surgery [5], this study mostly focus on 3 operating rooms namely general operation room, cardiovascular operation room and neurosurgery operation room. I.

Method There are three steps in this project as shown in Fig1. In the first step data such as basic information, Standards of surgery and clinical data were collected from the theatre to describe different procedures. Further to this, technical data were also obtained to describe advanced technologies and medical equipment which is used for each surgical operation. Next, categorization was done based on the type and complexity of a surgery. The data that is collected are from different sources such as clinical websites, companies and textbook. The second step, all the collected data are presented by the use of power point and converted into a program. In the third step, after finalizing, this application was given to a number of

ISBN: 978-1-941968-26-0 ©2015 SDIWC

Clinical:

In this section, the most common surgeries such as: endocrine surgery, breast surgery and so on were studied. In addition, there is more emphasis on more sophisticated surgeries such as heart valve disease, biopsy, craniotomy and etc. II.

Technical:

In this part, medical devices have been divided into 2 categories. In general surgery, we have referred to the names of all the equipment needed to perform flawless operations in the theatre. For each piece of equipment information such as different types of the equipment, faults and troubleshooting, block diagram, various parts of the equipment, and the operation have been

206


added. The application and usage of each piece of equipment has been independently discussed. The equipment referred to in this study includes Anesthesia machine [6] [7], Electro cautery [8], surgical light and etc. In cardiovascular surgeries apart from the set of equipment in a general surgery theatre, all the specialized equipment such as heart-lung machine [9], advance image guided, cerebral oximetry[10], and so on have been pointed out. Step 2 The language used to write the program is C#, which has been used to interpret the data into an application. Visual Studio supports the ability to create different dialog boxes from a form or a different dialog box, the ability to make calculations, and creating different forms all while maintaining the true performance of the GUI.

Step 3 The program was given to 3 groups of biomedical engineers to work with the technical part, healthcare clinicians to work with clinical part of program, and biomedical engineering students to work with both parts of the application. They were requested to use this application for a period of time to collect their opinion based on user friendly and complexity. They were also asked to comment whether there is a workable link between the clinical and the technical parts of the application. Result When the program was run as shown in Fig2. The program starts with the three main surgical categories of general surgeries, cardiovascular surgeries, and neurosurgeries. (Fig.2)

Windows Forms provide the project with components, such as dialog boxes, menus, buttons, and many other controls, that make up a standard Windows application user interface (UI). Also, the designer view in Visual C# Express Edition enables us to drag the controls onto our application's main form and adjust their size and position. This section of the code below handles an important piece of the interface; it directly connects the user interface to the functionality of the program. This function handles what happens when something on the form is clicked.

Fig.2-Three main surgical procedures A Click on each button gives the user the two options of entering into either clinical or technical sections. (Fig.3)

Private void button_Click (object sender, EventArgs e) { FormSuction f1 = new FormSuction(); f1.show(); this.Hide();

}

ISBN: 978-1-941968-26-0 ©2015 SDIWC

207


Fig.3-Clinical and technical sections

Fig.5-Troubleshooting part of anesthesia machine.

Clinical section contains information about various diseases and the surgeries associated with them. (Fig.4)

Fig.6-Component part of the anesthesia machine

Fig.4-An example of a general surgery Also the technical section supplies information about the types of surgical equipment and the table of troubleshooting (Fig.5), how to use the equipment and its block diagram and components of the equipment (Fig6).

The findings from the survey indicates that the majority of biomedical engineering students know about medical surgeries and devices related to that ;however, less than 50 percent of the respondents have information about devices used in neurosurgery and cardiovascular surgery and the diseases associated with them . Also clinicians and biomedical engineers mostly were satisfied with this application. Shortcoming This study was only based on the simulation of operating room and it has potential to upgrade according to new technology and devices. For future study simulation of NICU [11] and simulation of

ISBN: 978-1-941968-26-0 ©2015 SDIWC

208


Radiology can be added to this application to design and make a comprehensive hospital simulator (Fig.7). Also a graphical design for medical equipment needs to be added to provide 3D images, which allows user to view the product in 360’ degree review.

of the global volume of surgery: a modelling strategy based on available data. Lancet, 372(9633) [3] Null G, Dean C, Feldman M, Rasio D, Smith D. (August 2006), Death by medicine. Life Extension Magazine [4] Ramachandran, M. Gladman, M.A (09 Dec 2010).Clinical Cases and OSCEs in Surgery, London. United Kingdom: Elsevier Health. [5] RN, M. A. (2007). Pocket Guide to the Operating Room: F.A. Davis Company [6] Sinclair CM, Thadsad MK, Barker I. (2006).Modern anaesthesia machines. Contin Educ Anaesth Crit Care Pain, 75-8

Fig.7- Future study part Conclusion A user friendly simulation with essential information of both clinical and technical part of operation room, known as the heart of hospital, is an innovative idea. This simulator can be used to train nurses and technicians in order to be familiar with advanced technology and increase patient safety. Also, this program is a knowledge resource for biomedical engineering students and biomedical engineers to learn about different components of medical equipment, trouble shooting, and their functions. Finally, the cutting edge of this application is that it can be easily upgraded to cope with the new technology. In fact this program enables user to have comprehensive and multidimensional information about operation room.

[7] Hartle A, Anderson E, Bythell V, Gemmell L, Jones H. (2012). Association of anaesthetists of Great Britain and Ireland. Anaesthesia, 67:660-8 [8] M. Saaiq, S. Zaib, and S. Ahmad. (31 Dec 2012). Electrocautery burns: experience with three cases and review of literature. Ann Burns Fire Disasters, 203–206. [9] By About.com Inventors. "John Heysham Gibbon Heart Lung Machine – Pump Oxygenator." Retrieved December, 5, 2015 from [10] Murkin JM, Admas SJ, Novick RJ (2007). Monitoring brain oxygen saturation during coronary bypass surgery: a randomized prospective study. Anesth Aalag, 51-58 [11] Fanaroff AA, Martin J. (2006).Neonatal Perinatal Medicine: Diseases of the Fetus and Infant, 8th ed.Mosby, 791-804.

References [1] Henriksen K, Battles JB, Keyes MA, Grady ML, editors. (August 2008). Advances in patient safety: New directions and alternative approaches. AHRQ Publication [2] Weiser TG, Regenbogen SE, Thompson KD, Haynes AB, Lipsitz SR, Berry WR.(12 Jul 2008) An estimation

ISBN: 978-1-941968-26-0 ©2015 SDIWC

209