Enhancing Matrix Factorization-based Recommender Systems via Graph Neural Networks

Due to the serious information overload problem caused by the rapid development of the Internet, recommender system (RS) has been one of the most concerned technologies in the past decade. Accompanied with the prevalence of social networks, social information is usually introduced into RS to pursue higher recommendation efficiency, yielding the research of social recommendations (SoR). Almost all of existing researches of SoR just consider the influence of social relationships, yet ignoring the fact that correlations exist among item attributes and will certainly influence social choices. Therefore, this work introduces the graph neural networks to enhance matrix factorization-based recommender systems. And the proposal in this work is named GNN-MF for short. The user subspace and item subspace in matrix factorization are represented with the use of deep neural networks, in which parameters are learned by back propagation. The experiments well prove efficiency of the GNN-MF.