Community-Based Matrix Factorization Model for Recommendation

Although matrix factorization has been proven to be an effective recommendation method, its accuracy is affected by the sparsity of the matrix and it cannot resolve the cold start problem. Social recommendation methods have attracted much attention in solving these problems. In this paper, we focus on community discovery rather than individuals’ relations in the social network and propose a community-based matrix factorization (CommMF) model. It consists of two parts. One is a community detection algorithm Coo-game, proposed in our previous work and used here to divide the social network of users into multiple overlapping communities. It is based on the game theory and can fast detect overlapping communities. Since the users in the same community share the common interests such as scoring information, some of the null values in the scoring matrix can be filled according to the communities. This will help alleviate the sparsity of the scoring matrix and the cold start problem of new users. The other part is the matrix factorization model, which is used to recommend items to users. The model is trained by a stochastic gradient descent algorithm. The experimental results on real and simulated datasets show that CommMF can get higher accuracy with the help of community information compared with PMF and SocialMF models.