Learning to Denoise Unreliable Interactions for Graph Collaborative Filtering

Abstract

Recently, graph neural networks (GNN) have been successfully applied to recommender systems as an effective collaborative filtering (CF) approach. However, existing GNN-based CF models suffer from noisy user-item interaction data, which seriously affects the effectiveness and robustness in real-world applications. Although there have been several studies on data denoising in recommender systems, they either neglect direct intervention of noisy interaction in the message-propagation of GNN, or fail to preserve the diversity of recommendation when denoising. To tackle the above issues, this paper presents a novel GNN-based CF model, named Robust Graph Collaborative Filtering (RGCF), to denoise unreliable interactions for recommendation. Specifically, RGCF consists of a graph denoising module and a diversity preserving module. The graph denoising module is designed for reducing the impact of noisy interactions on the representation learning of GNN, by adopting both a hard denoising strategy (i.e., discarding interactions that are confidently estimated as noise) and a soft denoising strategy (i.e., assigning reliability weights for each remaining interaction). In the diversity preserving module, we build up a diversity augmented graph and propose an auxiliary self-supervised task based on mutual information maximization (MIM) for enhancing the denoised representation and preserving the diversity of recommendation. These two modules are integrated in a multi-task learning manner that jointly improves the recommendation performance. We conduct extensive experiments on three real-world datasets and three synthesized datasets. Experiment results show that RGCF is more robust against noisy interactions and achieves significant improvement compared with baseline models.