Social Collaborative Filtering Methods Based On Matrix Factorization

Recommender system is a wonderful and effective tool of handling “InformationOverload”. To accurately and actively provide users with their potentially interestedinformation or services is the main task of a recommender system. Collaborative filtering isone of the most widely adopted recommender algorithms, whereas it is suffering the issuesof data sparsity and cold start that will severely degrade the quality of recommendations.One potential way to solve these problems is by exploring available social networks.With the rapid development of web2.0technologies, in addition to the ratings of itemscontributed by users, the social information of users get much more readily obtained thanbefore through social networking services. It is believed that human beings usually acquireand disseminate information through their acquaintances such as friends, colleagues orpartners, which implies that the underlying social networks of users might play afundamental role in helping them filter information. Trust relationship is one of the mostimportant types of social information in that we are more likely to accept viewpoints fromwhom we trust. Thus, it has become a big opportunity as well as a big challenge to improverecommendation quality by sufficiently and effectively utilizing available trust information.This article proposes a novel social collaborative filtering strategy, trying to improvethe performance of collaborative filtering by means of elaborately integrating twofold sparseinformation, the conventional rating data given by users and the social trust network amongthe same users. Based on the basic idea, a general matrix factorization method has beenproposed to map users into low-dimensional latent feature spaces in terms of their trustrelationship, aiming to reflect users’ reciprocal influence on their own interest morereasonably and to learn more realistic interest patterns of users for better recommendations.The validations against four different real-world datasets show that the proposed methodperforms much better than the state-of-the-art recommendation algorithms for socialcollaborative filtering.