A research team has introduced a contrastive learning framework called Geometry-aware radial-angular perturbation and Topology-aware contrastive learning, or Geo-Top, designed to improve representation learning across heterogeneous information networks. Heterogeneous information networks model complex systems by integrating multiple node types and distinct semantic relations. However, existing heterogeneous graph contrastive learning methods face two persistent operational constraints. Standard random perturbations frequently disturb fragile cross-type structures, and conventional local message-passing encoders struggle to capture long-range dependencies alongside broader structural cues. To overcome these barriers, the research team developed Geo-Top to combine topology-aware structural encoding with geometry-aware radial-angular latent perturbation.
At the core of the framework is an SVD-inspired learnable low-rank topology factorization mechanism developed to model meta-path-induced adjacency relations. By decomposing these complex relational pathways into low-rank structural factors, the system injects the resulting topological properties directly into attention computation using gradient decoupling. This gradient decoupling design enables the model's topology-aware attention mechanism to couple local semantic information with non-local structural cues.
In parallel with its structural factorization module, Geo-Top addresses data augmentation limitations by introducing structured perturbations into the latent tangent representation. Instead of applying arbitrary stochastic noise that could corrupt sensitive cross-type network links, the framework implements geometry-aware radial-angular perturbations. This geometric strategy generates structured augmented samples. The complete framework integrates this latent perturbation method alongside shallow feature masking, a cross-view contrastive learning architecture, and momentum target encoders. Operating together, these modular components preserve semantic-view consistency and support effective topology-aware representation learning across contrasting views of the graph data.
The researchers evaluated Geo-Top across five public heterogeneous graph benchmark datasets: DBLP, AMiner, Freebase, ACM, and IMDb. In graph clustering evaluations, Geo-Top established leading performance on the reported clustering metrics for the DBLP, AMiner, and Freebase datasets. However, the experimental outcomes also revealed that clustering performance gains were not completely uniform, as the observed clustering benefits on the ACM and IMDb benchmarks depended on the specific characteristics of the dataset under evaluation.
To isolate the operational influence of each system module, the research team conducted systematic component ablations across both classification and clustering tasks. These ablation experiments characterized the specific empirical contributions made by the topology-aware encoding module and the latent radial-angular augmentation mechanism.
Subject of Research: Contrastive representation learning for heterogeneous information networks
Article Title: Integrating geometry-aware radial-angular perturbation and topology-aware contrastive learning for heterogeneous graphs
Article References: Sun, M., Luo, X., Hu, M., Huang, E., & Jiang, X. (2026). Integrating geometry-aware radial-angular perturbation and topology-aware contrastive learning for heterogeneous graphs. Scientific Reports. https://doi.org/10.1038/s41598-026-77826-6
Image Credits: AI Generated
DOI: 10.1038/s41598-026-77826-6
Keywords: heterogeneous information networks, contrastive learning, graph neural networks, topology-aware encoding, radial-angular perturbation, latent tangent representation, low-rank topology factorization, meta-path relations, gradient decoupling, momentum target encoders, feature masking, node classification, graph clustering, DBLP dataset, AMiner dataset, Freebase dataset, ACM dataset, IMDb dataset
Cite Scienmag News
Reid Dalton. (November 19, 2026). Geo-Top Framework Shows Strong Node Classification on Heterogeneous Graph Datasets. Scienmag. https://scienmag.com/geo-top-framework-shows-strong-node-classification-on-heterogeneous-graph-datasets/
Reid Dalton. "Geo-Top Framework Shows Strong Node Classification on Heterogeneous Graph Datasets." Scienmag, 19 November 2026, https://scienmag.com/geo-top-framework-shows-strong-node-classification-on-heterogeneous-graph-datasets/. Accessed 19 November 2026.
Reid Dalton. "Geo-Top Framework Shows Strong Node Classification on Heterogeneous Graph Datasets." Scienmag. November 19, 2026. https://scienmag.com/geo-top-framework-shows-strong-node-classification-on-heterogeneous-graph-datasets/








