PERSONALIZED RECOMMENDER SYSTEM COMBINING USER INTEREST AND SOCIAL CIRCLE

In this abstract, three social factors, personal interest, inter-personal interest similarity, and interpersonal influence, fuse into a unified personalized recommendation model based on probabilistic matrix factorization. The personality is denoted by user-item relevance of user interest to the topic of item. To embody the effect of user’s personality, we mine the topic of item based on the natural item category tags of rating datasets. Thus, each item is denoted by a category distribution or topic distribution vector, which can reflect the characteristic of the rating datasets. Moreover, we get user interest based on his/her rating behavior. We then assign to the effect of user’s personality in our personalized recommendation model proportional to their expertise levels. On the other hand, the user-user relationship of social network contains two factors: interpersonal influence and interpersonal interest similarity. We apply the inferred trust circle of Circle-based Recommendation (CircleCon) model to enforce the factor of interpersonal influence. Similarly, for the interpersonal interest similarity, we infer interest circle to enhance the intrinsic link of user latent feature. 1) Propose a personalized recommendation system combining user personal interest, interpersonal interest similarity, and interpersonal influence. The factor of user personal interest makes direct connections between user and item latent feature vectors. And the two other social factors make connections between user and his/her friends’ latent feature vectors. 2) Propose a personalized recommendation approach by enforcing user personal interests, which is category related and represented by a multi-level tree structure. Personal unique interest is modeled to get an accurate model for the cold start user and user with very few friends and rated items. The impacts of the three factors to the recommendation performances are systematically compared. 3) Extensive experiments based on three datasets including Yelp, MovieLens, and Douban Movie show the effect of proposed model to solve the user cold start and sparsity problem. 4) We share our datasets for researchers in social recommendation area. The most salient feature of the shared datasets is that objective social recommendation performance evaluation can be carried out.