Aiming at the universal problem of data sparsity and cold start in recommendation system, a social network recommendation method based on the combination of social label and trust relation is proposed in the paper. The proposed method uses the probabilistic factorization technique to integrate the social trust, item marking information and user item score matrix. The interconnection of users and potential item feature space is realized from different dimensions. On this basis, the realization of dimension reduction by the factorization of probability matrix can achieve effective social recommendation. Experimental results on Epinions and Movielens data sets show that the proposed method is superior to traditional social recommendation and social label recommendation algorithms, especially in the case of less user score data.