Abstract: Collaborative filtering (CF) recommender system has been a most successful recommendation model in the past decade. However, the sparseness of user-item matrix and cold-tart problem still remain the challenges. The emergence of online social networking provides a great deal of social trust information for recommender systems, thus providing an opportunity to solve these problems. In this paper, based on matrix factorization collaborative filtering method, a model of integrating user trust information is proposed. This method uses trust information of users to amend the user latent factors and employs an auto-encoder to extract the initialization features of user and item latent feature vectors. And then a trust group detection algorithm is proposed for the trust relationship in the social network. Extensive experiments on real data sets show that the proposed model can not only effectively alleviate cold start, but also achieve better recommendation performance than the compared algorithms.