User Trust Relation Based On Subjective-Qualitative Bayesian Method

As with the development and popularization of the Internet, application of social networks(SN) becomes more and more widely. Users in social networks accept recommendation and choose their friends with their knowledge of user trust information. Therefore, trust prediction in social networks has gained much interest in recent years, and research on mining and using trust information lying under the social networks has shown its importance.Currently, trust prediction methods focus on user profiles, community features and communication theory to fulfill their aim of predict trust relation between users. Classifier-based methods are widely used and improved with their effective consideration about user profiles, offering prediction for given pairs of users. However, intercommunication along paths formed by multiple users are ignored in such methods, sometimes resulting in over-fitting. On the other hand, methods using community knowledge may be rough on quantitive features because of their concentration on the structure of communities.There is another way of handling such prediction problem: using Bayesian Networks(BN). Bayesian Networks are quite complete in terms of foundation, being capable to reason with rich quantitive information. However, main disadvantage of Bayesian Networks-- its high complexity in both training and reasoning-- could not be ignored. To reduce such complexity, Qualitative Bayesian Networks(QBN) are introduced, with much lower complexity and highly abstracted features, and, at the same time, ambiguous reasoning results, which would usually not be produced by classical Bayesian Networks. Often, such ambiguities are found just conflicts in the reasoning procedure, and considered totally useless.This thesis presents Subjective Semi-Qualitative Reasoning Networks(SSQRN), based on theories of Qualitative Bayesian Networks and Subjective Bayesian Method. Subjective Semi-Qualitative Reasoning Networks are uncertainty reasoning networks with the basic structure of Qualitative Bayesian Networks and qualitative influences along with quantitive features. Such reasoning networks support evaluation and reasoning direction of both sufficiency and necessity, taking classical Qualitative Bayesian Networks as a trivial case when applied with extreme parameters. Corresponding to the laws of classical Qualitative Bayesian Networks, symmetry, transitivity and composition laws are presented to enable Subjective Semi-Qualitative Reasoning Networks to use some of classical qualitative reasoning algorithms. A novel user trust prediction method instance on Epinions web site is thus presented based on SSQRN theory. Test results shows it wins the baseline(Subjective-Bayesian-based) method, and some classifier-based methods in precision of prediction.