Cross-Branch Contrastive Learning Algorithm in Heterogeneous Networks

Abstract

Heterogeneous information networks (HINs) provide a flexible graph representation for relational data with different entity and relation types. Such heterogeneity enriches semantic modeling, but also increases the difficulty of representation learning. Existing heterogeneous graph neural networks have achieved promising results on HIN data; however, many of them rely on supervised or semi-supervised training and require sufficient labeled nodes. Since reliable annotations are costly in large-scale networks, reducing label dependence remains an important issue. To address this problem, this paper develops Cross-Branch Contrastive Learning (CBCL), a self-supervised framework for HIN representation learning. CBCL learns node embeddings through two complementary branches. The structural-view branch models heterogeneous neighborhood relations and path-induced structural dependencies, while the path-semantic branch captures meta-path-level semantic information. A cross-branch contrastive objective aligns the two representations and transfers complementary information between them. Moreover, meta-path-based relevance is used to expand positive samples, allowing semantically related nodes to participate in contrastive optimization. By integrating structural modeling, semantic fusion, and cross-branch supervision, CBCL learns discriminative node representations without extensive labeled data.