A Novel Neuro-Fuzzy System for Classification

Global Journal of Pure and Applied Mathematics (GJPAM) ISSN 0973-1768 Volume 12,Number 1 (2016) © Research India Publications : http://www.ripublication.com

A Novel Neuro-Fuzzy System for Classification S. Leoni Sharmila

C.Dharuman

P.Venkatesan

Asst.Professor, Department of Mathematics, SRM University, Ramapuram, Chennai, India. E-mail: [email protected]

Professor, Department of Mathematics, SRM University, Ramapuram, Chennai, India. E-mail: [email protected]

Professor, Faculty of Research, Sri Ramachandra University, Chennai, India. E-mail: [email protected]

logic and its relation with probabilistic logic is explored. [6] Classifies Renal Failure Using Simplified Fuzzy Adaptive Resonance Theory Map. In [7] a new brain-computer interface design has been created using fuzzy ARTMAP.

Abstract— One of the major challenges in medical diagnosis for proper treatment is fast and accurate diagnosis of the disease. Neural Networks are used as powerful classifier for early detection. Fuzzy logic which operates with many truth variables is applied to study uncertain, vagueness and complexity. The aim of this paper is to study the usefulness of fuzzy and neuro- fuzzy approach for classification and prediction of problems. Keywords—Fuzzy classifications, classifications, Simplified Fuzzy ArtMap.

I.

Neuro

–

III. Fuzzy Classification In the field of diagnosis structure is found from the data obtained from observations. Finding the structure is the essence of classification, by finding the structure it would be able to classify the data according to similar patterns, attributes, features and other characteristics. Fuzzy partitioning is a methodology for generating fuzzy sets to represent the underlying data. Fuzzy partitioning techniques are classified into three categories: grid partitioning, tree partitioning, and scatter partitioning. In this three partitioning methods, grid partitioning is the most commonly used in practice, particularly in system control applications. In Grid partitioning, partition forms by dividing the input space into several fuzzy slices, which is specified by a membership function for each feature dimension.

Fuzzy

Introduction

Fuzzy logic is good classifier of data. It deals with natural language in addressing impreciseness or vagueness. It provides a way to arrive to a definite conclusion for noisy, ambiguous or missing input information for the real world problems [1]. Neural network is a learning algorithm which gives a black box effect and cannot deal with imprecise data. The need for hybridization is to overcome the weakness in one with the strength of the other. Integrated Neuro-fuzzy system combines the learning abilities of neural network and explanation abilities of fuzzy system. The aim of this paper is that Neuro-Fuzzy system predicts better than Fuzzy Logic in classification.

II.

IV.

Neuro-Fuzzy Classification

Many complex domains have many different variety of problems, which require different types of processing. The use of intelligent hybrid systems is rapidly growing with many successful applications in areas including medical diagnosis, process control, engineering design, financial trading and cognitive simulation. Fuzzy logic provides an inference mechanism under uncertainty conditions, computational neural networks is advantages, such as learning, adaptation, fault tolerance, parallelism and generalization. Fusion of neural network and Fuzzy logic (Neuro- Fuzzy) is very efficient [ 8, 9]. Neural networks tune the membership functions of fuzzy systems that are employed as decision-making systems for controlling equipment.

LITERATURE REVIEW

Classification systems have been used for disease diagnosis problem and for other clinical diagnosis problems. There have been several studies reported focusing on disease diagnosis using Fuzzy and Neural Network. These studies applied different methods to achieve high classification accuracies for dataset. In [2] considered a Multi-layer Perceptron (MLP) using back propagation algorithm to classify thyroid disease. In[3] gives an approach for recognizing breast cancer using Neuro- Fuzzy Inference technique. In [4] developed a Neuro- Fuzzy Expert system for classifying heart disease . In [5] gives usefulness of fuzzy

267


V.

Simplified Fuzzy ART Map (SFAM)

a = 1− a

SFAM model is a combination of fuzzy logic and adaptive resonance theory (ART) neural networks. It is supervised network that can perform incremental learning. The benefit of this network is to lower the training time compared to other NNs. FAM and SFAM networks have been widely used in different classification problems [7, 10, 11, 12, 13].

(1)

The internal compliment code input vector I is then of dimension 2d. I = (a,

a ) = (a1, a2,… ad , a 1, a 2,… a d)

(2)

The activation functions and matching functions are defined as

SFAM is a fast, online, supervised learning method which uses simple fuzzy rules for neuron activation and selection. The main advantage of using fuzzy rules is that it reduces computation necessary for learning and that it can learn each pattern with small number of iterations. SFAM begins having no connection weight, it enlarges in size to fit the task and employs simple learning equations, and also contains just parameter selected by user recognized as vigilance parameter [6].

Tj =

M=

I ∧wj

(3)

α + wj I ∧W j

(4)

I

where W j are current values of templates (weight vector) SFAM consists of two layers: an input and an output layer. Block diagram of the SFAM network, the main architecture is shown in fig.1. The Input, into the network must be normalized to a value from 0 to1. So a suitable normalization value has to be chosen so that no input will fall outside the valid range. Input is been normalized by compliment coder and also provides the fuzzy compliment for each value. This input (I) is then passed to the input layer. Weights (w) from each of the output node sample the input layer, making the weights top-down. Training begins with just one hidden node and corresponding weights are set equal to the first record and prediction is set equal to the class of the first record. When a new class is encountered a new node is created. The node, whose weights match best to the current input, supplies the prediction provided, the degree of the match exceeds the vigilance threshold value. If this is a correct prediction, the weights of this winning node are adjusted to this input. If this a wrong prediction or vigilance threshold is not achieved, a new node is created with weights and prediction is equal to this record [6].

which is associated with output nodes j and α is a small value close to zero. The updates of templates that belong to the resonant domain is represented as an assignment statement.

W j = (1 − β )W j + β I ∧ W j

(5)

where β is the learning rate, 0≤ β ≤1. The operator

I ∧ W j = ∑ min(I ,W j ) used in (4) and (5) defines fuzzy AND function, which assumes positive, normalized values of the inputs [6].

Figure1- Architecture of SFAM

Network is said to achieve a state of resonance, if the Network function value exceeds vigilance parameter. Network is said to be mismatch, reset, if vigilance parameter exceeds match function value. Once the network is trained, by passing input pattern into coder and then input layer, all the output nodes compute activation function with respect to input. The winner is the node with the highest activation function, is chosen. The category of the input is found by assigning the category of the most highly activated node as max ( Tj ). The category layer holds the names of the (m) categories which the network is expected to classify [6].

VI.

Experimental Results and Discussions

A. Data set Description The training algorithm is now described for its completeness. For the given input vector a of d features, the compliment vector a represents the absence of each feature

The data set in this work is taken from UCI Machine Learning Repository. The data consists of thyroid disease classifies to 3 groups namely hypothyroidism,

268


hyperthyroidism and normal. Total numbers of samples are 215.The parameters are as follows:

Table 3 - Classification value of Neuro- Fuzzy for Test data

1. Class attribute (1 = normal, 2 = hyper, 3 = hypo) 2. T3-resin test. (A percentage) 3. Total Serum thyroxin is measured by the isotopic displacement method. 4. Total serum of triiodothyronine is measured by radio immuno assay. 5. basal thyroid-stimulating hormone (TSH) measured by Radio immunoassay. 6. Maximal absolute difference of TSH value after the injection of 200 micro grams of thyrotrophicreleasing hormone when compared to the basal value.

Predicted Value

Actual Value

Type class of the UCI classification 1

Hyper

30

2

Hypo

35

3

Total

215

44 0 0

3 7 0

1 0 10

Normal Hypo Hyper

Performances of the Fuzzy logic and neuro-fuzzy systems are summarized in Table 4. Results of other methods for thyroid classification are given in Table 5. Table 4 - Comparison of Fuzzy logic and Neuro – Fuzzy Neuro – Fuzzy systems

Fuzzy Logic

B. Experiment The Thyroid classification data set consists of 215 samples out this 150 patients does not have the disease. 30 patients have hyper thyroid and 35 have hypo thyroid. There are 6 inputs out of which the first attribute has to be classified. For fuzzy classification the tool used was Fuzzy Weka. In this Grid partition method was used with cross validation 10 folds. Unclassified instance is 1, For Neuro – Fuzzy classification, Neunet Pro SFAM method was used. In this experiment 70% split was used for training and 30% for testing, classification results are presented in table 2 and 3.

Normal

Hypo

Hyper

150 21 12

0 14 0

0 0 17

84%

94%

48%

100%

Specificity

100%

92%

Error Rate

15.4%

6.1%

Correct Classification

Predicted Value

Normal Hypo Hyper

Accuracy Sensitivity

Table 5 - Results of different methods

Table 2 - Classification value of Fuzzy logic

Actual Value

Hyper

The performances of the neuro-fuzzy classifier were evaluated using the following parameters: 1. Accuracy = TP+TN/ Total*100 is the correct classification rate. 2. Sensitivity= TP/ (TP+FN)*100 is the true positive rate. 3. Specificity = TN/ (TN+FP)*100 is the fraction of nonevents that has been correctly rejected. 4. Error rate = FP+FN/Total*100 is the error rate.

Table 1 - Distribution of patients and normals Number of Samples 150

Hypo

C. Results Analysis

All attributes are continuous. There is no missing value. The laboratory samples are determined based on the distribution of the type of disease [15].

Type disease Normal

Normal

Fuzzy Logic

84%

MLP[14]

90%

CSFNN[14]

83.9%

Neuro-Fuzzy

94%

V. Conclusion This paper presents a comparative study on thyroid disease diagnosis using Fuzzy Logic and Neuro- Fuzzy classification. From the experimental results it is noticed that Neuro-Fuzzy system gives 94% accuracy and Fuzzy logic model gives 84% accuracy of classifying thyroid diseases.The hybrid technique

269

Global Journal of Pure and Applied Mathematics (GJPAM) ISSN 0973-1768 Volume 12,Number 1 (2016) © Research India Publications : http://www.ripublication.com [14] C. Senol, T.Yildirim, “Thyroid and Breast Cancer Disease Diagnosis using Fuzzy-Neural Networks”, IEEE transcations on Electrical and Electronics Engineering, 2009, p. II 390 - II393.

could be successfully used to help the diagnosis of thyroid disease. The Neuro- Fuzzy system used in this study shows better performances than Fuzzy Logic. The advantage of SFAM is capable to perform classification very efficiently and giving very high performances.

[15] www.ics.uci.edu/pub/ml-repos/machine-learningdatabases/thyroiddisease

References [1] Lofti A. Zadeh, “Fuzzy logic, – the logic underlying approximate, rather than exact, modes of reasoning – is finding applications that range from process control to medical diagnosis”, niversity of California, Berkeley, 1971. [2] Gharehchopogh. F.S., Khalifelu.Z.A, “Neural Network Application in Diagnosis of Patient: A Case Study”, International Conference on Computer Networks and Information Technology (ICCNIT), pp. 245 – 249, 2011. [3] Fatima. B., Amine.C. M., A Neuro Fuzzy Inference model for breast cancer recognition, International Journal of Computer Science & Information Technology (IJCSIT), 2012. [4] Ephzibah E.P., Dr. Sundarapandian.V, “ A Neuro – fuzzy Expert system for heart disease diagnosis” , Computer Science & Engineering : An International Journal (CSEIJ), Vol.2, No.1, Feb 2012. [5] Licata.G, “ Employing fuzzy logic in the diagnosis of a clinical case”, Journal Health, Vol:2, No:3, pp. 211-224. [6] P. Venkatesan and M. L. Suresh, "Classification of Renal Failure Using Simplified Fuzzy Adaptive Resonance Theory Map," IJCSNS International Journal of Computer Science and Network Security, vol. 9, no. 11, pp. 129- 134, November 2009. [7] R. Palaniappan, P. Raveendran, S. Nishida and N.Saiwaki, "A new brain-computer interface design using fuzzy ARTMAP," IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 10, no. 3, p. 140–148, 2002. [8] Imianvan Anthony.A, Obi Jonathan.C, “ Application of Neuro – Fuzzy System for the Probe and Prognosis of thyroid disorder”, Disorder International Journal of Fuzzy Logic Systems ((IJFLS) Vol.2, No.2,April 2012. [9] Fatima. B., Amine.C. M., “A Neuro Fuzzy Inference model for breast cancer Recognition”, International Journal of Computer Science & Information Technology (IJCSIT), 2012. [10] G. A. Carpenter, S. Grossberg, N. Markuzon, J. H.Reynolds and D. B. Rosen, "Fuzzy ARTMAP: a neural network architecture for incremental supervised learning of analog multidimensional maps,"IEEE Transactions on NeuralNetworks, vol.3, no.5,pp. 698– 713, 1992. [11] T. Kasuba, "Simplified fuzzy ARTMAP," AI Expert,vol. 8, no. 11, p. 19–25, 1993. [12] I. Dagher, M. Georgiopoulos, G. L. Heileman and G.Bebis, "An ordering algorithm for pattern presentation in Fuzzy ARTMAP that tends to improve generalization performance," IEEE Transactions on Neural Networks, vol. 10, no. 4, p.768–778, 1999. [13] P. Raveendran, R. Palaniappan and S. Omatu, "Fuzzy ARTMAP classification of invariant features derived using angle of rotation from a neural network," Information Sciences: An International Journal, vol. 130, pp. 67-84, 2000.

270