The Recommendation Algorithm Of Trust-fused Matrix Factorization Based On Boosting

Collaborative filtering is most widely used. However, it has some problems in the cold start, sparsity, trust recommendation and the overfitting according to recent studies. With the continuous expansion of the scale of the data, these problems have become increasingly prominent. In view of the above problems, we put forward a kind of recommendation algorithm named the Trust-fused MF Based on Boosting.Firstly, through the analysis of the influence of trust relationship between users on the recommendation results, the Trust-fused Matrix Factorization based on the Biased MF model is put forward. The algorithm puts users’ explicit effects of trust (the trust value between users) and implicit effects of trust (users’ trusting friend) into the matrix factorization recommendation model.It can effectively solve the cold start and the sparsity problem by the fusion of explicit effects of trust, and at the same time, it can achieve higher recommendation accuracy by the fusion of implicit effects of trust.Secondly, in order to further improve the ranking performance, the Matrix Factorization Based on Boosting by the method of Integrated learning is designed. It is a sort of oriented Top-N recommendation algorithm. This framework can be expressed as a fusion of the AdaBoost algorithm and MF algorithm. At the same time, the ensemble framework uses NDCG in ranking as error evaluation index through linear integration of weak recommenders in order to obtain better performance.Finally, the Trust-fused Matrix Factorization is used as weak recommender to boost learning. The whole model is called the Trust-fused MF Based on Boosting which can optimize the recommendation list for ranking more directly and prevent easy overfitting shortcomings of the traditional MF effectively.In this paper, the datasets of Netflix, Epinions and MovieLens are used. The experimental analysis of the Trust-fused Matrix Factorization, the Matrix Factorization Based on Boosting and the Trust-fused MF Based on Boosting are made. Their experimental comparisons with some common recommendation algorithms are also done, and the effectiveness of the algorithms is verified.