Multi-view social recommendation via matrix factorization with sub-linear convergence rate

Abstract

Recently Recommender Systems (RS) have become crucial tools to deal with information retrieval and filtering in various applications, such as online business, social network and so on. However, with the ever-increasing scale of social network and information overload, RS still suffer from data sparsity and heavy computational complexity, and even confront a serious challenge to further improve the recommender performance and user experience. To this end, this article propose a novel multi-view social recommendation scenario referred to as MsRec, which takes account of available information for item recommendation from multiple perspectives. More specifically, MsRec tries to exploit intricate inner relationship within social network, and perform user-level preference learning for each user without overlapping, moreover, MsRec leverages available side information, such as contextual information, demographic characteristics and item attributes, to perform representation vector learning for items, which will be further utilized for item similarity calculation via Cosine method. With the learned social influence and item similarity coefficient, MsRec could perform representation vector learning for users and items, and further provide rating prediction and item recommendation. In addition, theoretical analysis indicates that MsRec could achieve convergence with sub-linear convergence rate during model training phase, and yield high performance in social recommendation, without heavy computational cost. Experimental analysis over Yelp, Epinions and DoubanMovie also demonstrate the superiority of the proposed MsRec, which can achieve significant improvements while compared to other benchmark recommender algorithms, and provide much better user experience.