Learning to Denoise Unreliable Interactions for Graph Collaborative Filtering

Changxin Tian, Yuexiang Xie, Yaliang Li, Nan Yang, Wayne Xin Zhao

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.