A Research Of Recommendation Based On Trust Inference And User Feedbacks

With the rapid development of Internet technology, there have appeared many kinds of social media which are based on user relationships. With the quickly bursting Internet information and the accelerated life rhythm, filtering redundant information and capturing users’ interests quickly have become one of the core appeals of many so-cial media. Recommender system can make use of many kinds of information, e.g. user feedbacks, user relationships, item contents and so on, to analyze users’ preference and filter information according to users’ preference so as to help users to get useful infor-mation quickly. The basic user feedback based recommendation, no matter multi-class recommendation or one-class recommendation all have poor performance because of sparse problem. Many researchers try to use user relationships to improve recommen-dation effect. Although this kind of recommender methods has achieved some effect, it is also facing many problems, such as how to deal with the sparse user relationships, how to make use of trust relationships to deal with multi-class recommendation and one-class recommendation respectively. To deal with these problems, we propose a method which firstly fills original trust relationship based on the observed trust, and then recommends with the more complete trust relationships. The main contributions are as follows:1. We propose a recommender system framework based on trust relationship infer-ence. The framework combines trust inference and feedback inference together, it can solve the problems caused by sparsity of user relationships and user feedbacks and deal with multi-class recommendation and one-class recommendation respec-tively.2. We propose a trust inference method which combines non-negative constraint and similarity with existing methods. This method extends existing multi-aspect trust inference model with non-negative constraint and similarity, and finally combines the extended models together by logistic regression.3. We propose a trust-based multi-class recommendation model based on matrix fac-torization and social regularization. This model first infers several types of unob-served trust based on the observed trust relationships, and incorporates observed trust and inferred trust with genetic algorithm, then generates the model using PMF methods with homophily effect.4. We propose a trust-based one-class recommendation model based on matrix factor-ization and social regularization. This model first infers several types of unobserved trust based on the observed trust relationships, and incorporates observed trust and inferred trust with NMF method for better recommendation.