Efficient GNN-based Social Recommender Systems through Social Graph Refinement

Abstract Precisely recommending relevant items to users is a challenging task because the user’s rating can be influenced by various features. Therefore, social recommender systems have recently been introduced to leverage both the user-item interaction graph and the user-user social relation graph for more accurate rating predictions. Moreover, as Graph Neural Networks (GNN) have demonstrated superior performance in graph representation learning, several algorithms have been developed to incorporate GNN into social recommender systems. However, when the sizes of the social graph and user-item graph are very large, the computational complexity of existing GNN-based social recommender systems for aggregating user and item nodes becomes the primary bottleneck. In this paper, we develop a novel lightweight GNN-based social recommender systems (called LiteGSR) that mitigates the model complexity associated with aggregation operations while maintaining the quality of convergence. In order to efficiently reduce computational complexity, our proposed model refines the social graph by utilizing PageRank-based centrality scores of users and adapts representative virtual users on the user-item graph. Experimental results demonstrate that our new social recommender systems outperform existing state-of-the-art recommender systems in both convergence and training time.