A recommender system fused with implicit social information through network representation learning

Abstract

Traditional recommender systems have always suffered from cold start and data sparsity problems by leveraging only a rating matrix. Social recommender systems try to address such problems by leveraging the user's social information, such as trust relationships, as auxiliary information. However, such explicit social information is not always available. In this paper, an implicit interest relationship mining approach based on network representation learning for recommendation, called INRec, is proposed to deeply mine implicit social information and fuse both explicit and implicit social information into a matrix factorization (MF) model and thus improve the effectiveness of the recommendations is proposed. Specifically, a modified one-mode projection method is first designed to construct an interest-connected user interaction network, and a network representation learning method is then employed to generate low-dimensional user representations. Abundant user interest relationships can be later obtained from user representations through a similarity calculation. Finally, a multi-source information fusion approach is applied to incorporate explicit and implicit social relationships into the MF model. The experimental results demonstrated that the average MAE and RMSE by INRec achieves 0.7415 and 0.9633 on three datasets (FilmTrust, CiaoDVD, Epinions), and it remarkably outperforms the state-of-the-art by a large margin.