Answers are grounded in this paper and cite their sources.
Heterogeneous information networks (HINs) model multiple node types and semantic relations, but existing heterogeneous graph contrastive learning methods still face two limitations. Random perturbations may disturb fragile cross-type structures, and local message-passing encoders may be limited in capturing long-range dependencies and broader structural cues. To address these limitations, we propose Geometry-aware radial-angular perturbation and Topology-aware contrastive learning (Geo-Top), a heterogeneous graph contrastive learning framework that combines topology-aware encoding with geometry-aware radial-angular latent perturbation. Geo-Top uses an SVD-inspired learnable low-rank topology factorization to model meta-path-induced adjacency relations and injects the resulting structural factors into attention computation through gradient decoupling. This design enables topology-aware attention to couple local semantics with non-local structural cues. Geo-Top further introduces radial-angular perturbations in the latent tangent representation to generate structured augmented samples. Together with shallow feature masking, cross-view contrastive learning, and momentum target encoders, these modules support semantic-view consistency and topology-aware representation learning. Experiments on five public heterogeneous graph datasets show strong node-classification performance. Geo-Top leads the reported clustering metrics on DBLP, AMiner, and Freebase, while the results on ACM and IMDb show that clustering benefits depend on the dataset. Component ablations characterize the contributions of topology-aware encoding and latent augmentation across classification and clustering.
No takes yet. Share an insight, caveat, or question.
Sun et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: