Simple Yet Effective Heterogeneous Graph Contrastive Learning for Recommendation

Contrastive learning has emerged as a prominent approach in Heterogeneous Graph Neural Network (HGNN)-based recommender systems, exhibiting particular efficacy in addressing the sparsity challenge inherent in user-item interaction data. However, these heterogeneous contrastive learning recommendation models still face two main limitations: (1) Contrastive learning methods typically generate contrastive views through node dropping, edge perturbation, and random augmentation techniques, which not only have high complexity but may also disrupt the graph structure and introduce noise. (2) Most HGNN-based contrastive recommenders are susceptible to the phenomenon of over-smoothing, leading to indistinguishable representation problem. In this study, we introduce SimHGCL, a novel and efficient heterogeneous graph contrastive learning recommendation model. Our approach innovates by generating only two matrices derived from meta-paths on the user-item interaction graph, thereby eliminating the need for computationally expensive node dropping and edge perturbation operations. By leveraging diverse matrices based on meta-paths, we facilitate the creation of distinct contrastive views, thereby substantially streamlining the contrastive view generation process. Moreover, to address the challenge of indistinguishable representation resulting from over-smoothing in contrastive recommenders, we propose a learnable Laplacian matrix exponent. Comprehensive empirical evaluations conducted on four real-world datasets demonstrate that SimHGCL significantly outperforms state-of-the-art baselines in terms of standard normalized discounted cumulative gain and recall metrics, while concurrently achieving substantial reductions in computational time. To facilitate the reproducibility of our work, we have open-sourced our code at https://github.com/zhanghaiyan1018/SimHGCL.