Research Of A Social Recommendation System Based On Topic Models

With the explosive growth of online social networks, the social network based approaches to recommendation have been well studied and developed. In the real world, users of social network always choose informations with respect to their content. But most social recommender systems are built on the combination of collaborative filtering and social trust propagation model, without concerning the text content of items. To increase the accuracy of social recommendation, we introduce topic models into social recommendation, taking the topics of items as factors of users’ratings.We first analyse the techniques of topic modeling and social recommending. Based on our studies, we propose a trust-based topic model named TB-LDA (Trust-Based factorization model of Latent Dirichlet Allocation) for recommendation systems. In model fitting procedure, we factorize the rating matrix to obtain latent feature factors, and consider the influence of social connections on personal affinity. Fitted model can be used to predict unknown ratings and provide recommendation for users. In order to form more predictive item factors, we adopt LDA topic model to infer latent topics of items. The training phase of LDA takes observed ratings as basis of posterior probability. This supervised learning method will achieve better predictions for parctical purpose of recommendation. Then we conduct experiments on a data set crawled from Weibo.com. The experimental results demonstrate that our approach performs better in accuracy of recommendation among social networks, despite data sparsity.