Recommender system, which helps users to acquire the information needed from enormous amounts of data, has becoming an effective information filtering technology. Matrix factorization is an efficient and effective method, and has received more and more attentions in the industrial community and academia since the competition of recommendation algorithm was organized by Netflix in October 2006. However, the existing matrix factorization models only take user behavior and social relationships into consideration, and ignore the fact that item relations also have an important influence on recommendation results in the social network-based recommender scenario. However, it is difficult to accuratly obtain item relations. Furthermore, when a system recommends items to a target user, how to consider both social relationships and item relations is an urgent issue.In response to the above problems, we firstly give a formula of measuring the related degree between items, based on this, we can mine item relations. Next, we proposed the dual regularized matrix factorization model, which integrates item relations-based related regularization and social relationships-based social regularization. The model can deal with the cold-start problem encountered by traditional matrix factorization models to some extent. Finally, we give a high accuracy algorithm: CRSVD++, which provides an effective method to solve rating prediction task. Extensive experimental results on four real datasets(Epinions,Flixster,Ciao and FilmTrust) show that, Compared to PMF and SVD++, CRSVD++ algorithm has obvious advantages; Compared to SoReg、SoRec、SocialMF、TrustMF, CRSVD++ algorithm can accurately predict the user’s real rating. |