Bayesian Networks to Predict Reputation in Virtual

Motiv. Def. Desi. Impl. Conclu. Biblio.

Bayesian Networks to Predict Reputation in Virtual Learning Communities Luis Chamba-Eras A, B Ana Arruarte and Jon Ander Elorriaga

B

Carrera de Ingenier´ıa en Sistemas - Universidad Nacional de Loja (UNL), Ecuador Computer Languages and Systems Department - University of the Basque Country (UPV/EHU), Ga-Lan Group, Spain ”2016 IEEE Latin American Conference on Computational Intelligence”

A B

Cartagena-Colombia: 2-4 November 2016

1 / 32


Outline

Motivations Definitions and related work Design of the Bayesian Network Implementation of prototype Conclusions and future work Bibliography

2 / 32


Outline


3 / 32


Motivations

• With the Internet Of Things and the support given by the In-

formation and Communication Technologies, real time participation and collaboration between individuals in different geographical locations is a reality in e-learning. • Currently there is great interest in predicting the indirect trust

or reputation among members of a Virtual Learning Communities (VLC): students and teachers. Any trust and reputation model have been implemented in the field of education, particularly in the VLC using aggregation algorithms, to estimate the capture of previous interactions of the members on resources or activities value reputation.

4 / 32


This is the reality! • ”I like” or ”I don’t like”

Figure 1: Sites that implement social computing 5 / 32


Outline


6 / 32


Preliminary concepts • Bayesian Networks (BN), known as probabilistic models or be-

lief networks, have been investigated due to a growing interest in predicting future events, a BN in general is a relationships network that uses statistical methods to represent probability relationships between different nodes. It is a compact representation of the joint probability distribution to reason under uncertainty. • Virtual Learning Communities (VLC), enabling members to pro-

duce knowledge, resulting from social interaction in a collaborative learning process. • Trust, concept is complex, so there are multiple definitions in

different contexts, initially, it was defined as the extent to which an individual has confidence and is willing to interact with someone based on words, actions and decisions of others. 7 / 32


Preliminary concepts

• Reputation, it is defined as the opinion that someone has about

something or somebody collected through indirect experiences. The reputation of a member of a VLC is the opinion that other members have on him. This opinion is based on the record of positive and negative interactions executed by them. • Positive reinforcement, is the reward offered to the member of

the VLC after performing a desired behavior, thereby determining the presence of this reward increases the probability that a behavior will occur. • Negative reinforcement, is the result offered to the member

after the appearance of unwanted conduct.

8 / 32


Preliminary concepts

• Online reputation, mid systems and unmediated. The mid sys-

tems, from simple systems are the spaces type of review of consumers using aggregation algorithms, for example as type consumer review sites: Yelp, Amazon, eBay, TripAdvisor, RealSelf.com, Menelaus, IMD, among others, or, to complex systems ratings as Moody. The unmediated systems are those in which the information provided by community members flows freely between all of them while mid systems needs a third agent that collects, stores, organizes and publishes. Examples of not mediated systems are recommendation letters such as those in LinkedIn, StackOverflow, reports Infojob, word of mouth networks such as Facebook or forums.

9 / 32


Related work

Related work combining BN, trust and reputation: • Trang Nguye et al. • L´ opez-Faican et al. • Daniel et al. • Qi et al. • Li et al. • Jøsang et al. • Patel et al. • Aciar et al.

10 / 32


Outline


11 / 32


Previous definitions

Figure 2: Acronyms of the context 12 / 32


Factors: direct experience and reputation

• Direct experience (F1): it is based on satisfaction concept and

it is a critical factor. Direct experience is obtained interacting with the ISs in the VLC. This is not always equal since in every society there are different points of view. An aggregation algorithm adapted to the VLC area, calculates the direct experience considering the interaction of members of the VLC with resources and learning activities managed in an LMS. Concretely, the algorithm considers the ”I like” actions (positive reinforcement) and ”I don’t like” actions (negative reinforcement) that each member performs on the resources/activities used and managed by the LMS.

13 / 32


Factors: direct experience and reputation

• Reputation (F2): in our proposal reputation factor (F2) is cal-

culated from past interactions achieved through direct experience (F1) between ISs in the VLC. The reputation of a member is the opinion that the other members have on him. This opinion is based on the history of positive and negative interactions carried out by them. People trust more in those individuals that have higher affinity. This factor is useful when there is little previous direct experience between the IS in the VLC.

14 / 32


Bayesian Network Model

Figure 3: BN factor by reputation of the TM 15 / 32


Formalization Bayesian Network Bayes Theorem:

Figure 4: Bayes Theorem

16 / 32


Formalization Bayesian Network Initial training values: estimates for positive/negative reputation

Figure 5: Training values BN 17 / 32


Formalization Bayesian Network Initial training values: positive reputation estimated values

Figure 6: Positive reputation estimated values 18 / 32


Formalization Bayesian Network Initial training values: negative reputation estimated values

Figure 7: Negative reputation estimated values 19 / 32


Formalization Bayesian Network

BN obtains for each member the probability of positive and negative reputation.

Figure 8: Estimation of positive and negative reputation for individual member in VLC

20 / 32


Outline


21 / 32


Implementation details in Moodle

• LMS Moodle 2.8.2. • Moodle Core, new modules and plugins. • Implement the direct experience and reputation. • Show resources: https://goo.gl/dOd6gO

22 / 32


Implementation details in Moodle The prototype enables two options ”I like” or ”I don’t like” on the LA and LR, which will be evaluated by each member based objectively on the contribution of these to their learning.

Figure 9: Implementation of ”I like” or ”I don’t like” in the forum activity by direct experience 23 / 32


Implementation details in Moodle When the member is on the VLC the ”trustmodel” block presents summary information of the scores that the member has achieved.

Figure 10: Trust data based on the TM for each member in the VLC by reputation

24 / 32


Using real scenario in VLC The prototype in real scenario for VLC, the process was as follows: 1. Start and management the virtual community for learning ”Mathematical Software” in Web. 2. 24 members enrolled in VLC, two teachers and 22 students. 3. Teachers tutoring and sharing activities/resources in VLC for two month. 4. Interaction and participate with members in VLC. 5. Recollected dataset about reputation factor in prototype using the ”trustmodel” block. 6. Comparison the dataset with the BN training values implement in BayesiaLab.

25 / 32


Using real scenario in VLC The values calculated by positive reputation compared to estimates reputation using the BN with the software BayesiaLab, identified that there are small differences between the values calculated with the prototype and BayesiaLab.

Figure 11: Comparative reputation result 26 / 32


Outline


27 / 32


Conclusions and future work • The main contributions of this paper are the development and

implement the BN for predicting reputation of members of VLC. • In this paper, we present the design of a BN that predicts repu-

tation from past interactions achieved through direct experience between members of a VLC. It has been implemented in the Moodle LMS. • In the work presented here the eBay aggregation algorithm

adapted to the VLC area has been implemented. It calculates the direct experience considering the interaction of members of the VLC with resources and learning activities managed in the Moodle. The algorithm considers the behavioral Psychology with ”I like” actions as positive reinforcement and ”I don’t like” actions as negative reinforcement.

28 / 32


Conclusions and future work

• In our work the past interaction is gathered from direct experi-

ence factor and use to predict value reputation factor. • As future work, consider the views of members with different

features present in a VLC and combining the BN with other Artificial Intelligence techniques as Natural Language Processing to identify through forums reputation.

29 / 32


Outline


30 / 32


Bibliography • Aguilar, J. (2015). Confianza y reputaci´ on en Sistemas Multi-Agentes. Universidad de los Andes. • Chamba-Eras, L. (2011). Modelo de Confianza para Objetos de Aprendizaje en Comunidades Virtuales. Master Thesis. Universidad del Pa´ıs Vasco. • Esfandiari, B. and Chandrasekharan, S. (2001). On how agents make friends: Mechanisms for trust acquisition. In Proceedings of the Fourth Workshop on Deception, Fraud and Trust in Agent Societies, pages 27-34, Montreal, Canada. • Gambetta, D. (1990). Trust: Making and Breaking Cooperative Relations, chapter Can We Trust Trust?, pages 213-237. Basil Blackwell, Oxford. • Garc´ıa, A. (Ed.), Ruiz, C. y Dom´ınguez, F. (2007). De la educaci´on a distancia a la educaci´ on virtual. Barcelona: Ariel. 31 / 32


(Creative Commons) BY-NC-SA

Thanks, Researchers: Luis Chamba-Eras, Ana Arruarte and Jon A. Elorriaga email: [email protected], [email protected], [email protected] Research Group: Ga-Lan, http://galan.ehu.es Twitter: @lachamba

32 / 32