Hierarchical Cross-view Contrastive Learning for Heterogeneous Graph Neural Network

Heterogeneous Graph Neural Networks (HGNNs) are widely used for modeling and analyzing heterogeneous information networks (HINs). However, most existing HGNN meth-ods are supervised or semi-supervised, hindered by the sparse supervised signal problem. Contrastive learning (CL) has emerged to address this issue by uncovering hidden infor-mation within data. Nonetheless, current CL methods often fail to harness the potential for extracting valuable information from similar nodes, either within a single view or across multiple views. This paper proposes a novel CL framework aimed at improving the quality of node embeddings on HINs, called Hierarchical Cross-view Contrastive Learning (HCCL) for heterogeneous graph neural networks. The framework builds en-hanced views from structural and semantic perspectives: the structural view captures extrinsic information by constructing a subgraph for each node, while the semantic view captures intrinsic information by leveraging meta-paths. Contrastive learning is applied both within each view and across views to obtain richer node representations. Com-prehensive experiments on four public datasets validate the effectiveness of HCCL. The source codes and datasets are available at https://github.com/lostraum/HCCL .