SIHAN: Semantic-Instance Guided Hypergraph Attention Network With Dual-View Contrastive Learning for Heterogeneous Graph

Abstract

Heterogeneous graphs are well-suited to modeling the diverse types of entities and their complex interactions in the real world. However, existing Heterogeneous Graph Neural Networks (HGNNs) are typically based on the binary message-passing framework, which struggles to explicitly and finely describe the higher-order semantic information encoded in meta-path instances. Recent studies have attempted to incorporate hypergraphs to capture higher-order interactions, yet most rely on co-occurrence aggregation of meta-path patterns, leading to the blending of instances with different semantics. Moreover, current hypergraph attention mechanisms fail to accurately differentiate the varying contributions of different semantic relations to node representation learning. To address these issues, we propose a novel fine-grained higher-order semantic modeling scheme based on meta-path instances, named SIHAN. First, we treat the complete sequence of intermediate nodes in each path instance as the semantic unit and construct a semantic-driven heterogeneous hypergraph, thereby structurally preserving instance-level semantics. Second, we develop a semantic-sensitive hypergraph attention mechanism that quantifies the contributions of different semantics to target node representation learning by measuring the semantic similarity between nodes and semantic hyperedges. Furthermore, to jointly model higher-order semantics and local topological structures, we construct a semantic-instance view and a structural-neighborhood view, and employ dual-view contrastive learning to encourage the model to simultaneously capture global semantic correlations and local structural regularities. Extensive experiments conducted on four real datasets demonstrate that our proposed model outperforms existing popular models across multiple downstream tasks.