HetDualCL: Dual-encoder contrastive learning for heterogeneous graphs

Heterogeneous graphs (HGs) model complex systems with multiple node and relation types, but their representation learning often relies heavily on labeled data and suffers from limited semantic integration. Current self-supervised methods are often limited to a single information perspective and encoder architecture, which hinders the full integration of multi-granularity semantic information. To bridge this gap, we propose HetDualCL, a novel dual-encoder contrastive learning framework that systematically integrates local topology from the network schema view and long-range semantics from the multi-hop view. We design an enhanced GNN encoder with normalization, residual connections, and feed-forward networks for robust local modeling, and a Gated Causal Convolutional (GCC) encoder to capture multi-hop semantic dependencies. A cross-view contrastive loss aligns and optimizes the two views collaboratively. Extensive experiments on four benchmarks show that HetDualCL learns highly discriminative node representations and achieves superior performance in node classification and clustering tasks.