Community-aware graph contrastive learning for collaborative filtering

Dexuan Lin

doi:10.1007/s10489-023-04787-y

2023-08-12

PRE

DOI：10.1007/s10489-023-04787-y

原文链接

原文求助

摘要

En 中文

Recently, graph neural networks have demonstrated superior performance in the field of collaborative filtering (CF). The graph collaborative filtering (GCF) method learns the interactions between users and items, whose performance is susceptible to sparse data. Self-supervised contrastive learning, which provides additional pseudo-labels besides the interaction data, has been widely adopted by existing GCF methods to solve the data sparsity problem. Despite their success, most existing methods destroy the semantics and ignore the community structure of the graph, thus failing to exploit the full potential of contrastive learning. In this work, we propose a novel graph contrastive learning paradigm called Community-Aware Graph Contrastive Learning (CA-GCL), which explicitly incorporates the community structure into contrastive learning to gain recommendations. Specifically, instead of random graph augmentation, we perform graph partitioning on the user-item bipartite graph in the pre-processing stage and construct the contrastive pairs based on the partitioning results, which in turn facilitates the representation learning of the model to gain downstream recommendation tasks. In addition, we add virtual nodes to the original graph based on the communities to accelerate the convergence of the model. Compared with the existing work, CA-GCL directly learns through the community structure of the graph, which ensures the high efficiency of the model while maintaining semantic integrity. Extensive experiments are conducted on three benchmark datasets, and the results show that CA-GCL significantly outperforms other graph collaborative filtering methods, demonstrating the feasibility and efficiency of CA-GCL.

Keyword:

Graph neural networks

Graph contrastive learning

Recommender systems

Collaborative filtering