Multi-view self-supervised learning on heterogeneous graphs for recommendation

Graph neural networks (GNNs) have significantly contributed to data mining but face challenges due to sparse graph data and lack of labels. Typically, GNNs rely on simple feature aggregation to leverage unlabeled information, neglecting the richness inherent in unlabeled data within graphs. Graph self-supervised learning methods effectively capitalize on unlabeled information. Nevertheless, most existing graph self-supervised learning methods focus on homogeneous graphs, ignoring the heterogeneity of graphs and mainly considering the graph structure from a single perspective. These methods cannot fully capture the complex semantics and correlations in heterogeneous graphs. It is challenging to design self-supervised learning tasks that can fully capture and represent complex relationships in heterogeneous graphs. In order to address the above problems, we investigate the problem of self-supervised HGNN and propose a new self-supervised learning mechanism for HGNN called Multi-view Self-supervised Learning on Heterogeneous Graphs for Recommendation (MSRec). We introduce a maximum entropy path sampler to help sample meta-paths containing structural context. Encoding information from diverse views defined by various meta-paths, decoding it into a semantic space different from own and optimizing tasks in both local-view and global-view contrastive learning, which facilitates collaborative and mutually supervisory interactions between the two views, leveraging unlabeled information for node embedding learning effectively. According to experimental results, our method demonstrates an optimal performance improvement of approximately 7% in NDCG@10 and about 8% in Prec@10 compared to state-of-the-art models. The experimental results on three real-world datasets demonstrate the superior performance of MSRec compared to state-of-the-art recommendation methods.