Research map / Graph learning / GNN architectures
Graph Neural Networks with Soft Association between Topology and Attribute
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GNN-SATA, a GNN that learns separate attribute and topology embeddings linked by a soft association term, with learnable edge pruning (GPM) and augmentation (GAM) modules, for homophilic and heterophilic node classification.
Problem
Interference between graph topology and node attributes distorts node representations, and most GNNs rely on homophily assumptions, so they perform poorly on heterophilic graphs.
Approach
Learns attribute (U) and topology (Z) representations separately, couples them through a soft association term that fits U U^T to the adjacency, and adaptively prunes existing edges and adds missing edges via learnable matrices (GPM and GAM) so the graph structure better matches homophily or heterophily.
Key results
- GNN-SATA reaches 65.25% accuracy on Squirrel, a 2.81% gain over the best baseline, while keeping only 27.19% of the original adjacency edges.
- On average GNN-SATA improves about 1.82% over the best baseline across six datasets; it reaches 91.24% on Cora and 77.12% on Chameleon.
- Ablation shows removing GPM or GAM or both lowers accuracy, e.g. Squirrel drops from 65.25 to 61.23 without both modules (SATA-AP).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add soft association between attribute and topology (ablation) | GNN-SATA · Squirrel | accuracy: 35.23 -> 65.25 | Helped |
| add GPM edge pruning (ablation) | GNN-SATA · Squirrel | accuracy: 61.95 -> 65.25 | Helped |
| add GAM edge augmentation (ablation) | GNN-SATA · Squirrel | accuracy: 63.38 -> 65.25 | Helped |
| add both GAM edge augmentation and GPM edge pruning (ablation) | GNN-SATA · Squirrel | accuracy: 61.23 -> 65.25 | Helped |
| use GNN-SATA instead of Ordered GNN | node classification · Squirrel | accuracy: 62.44± 1.96 -> 65.25 ± 1.33 | Helped |
Surprising findings
- Removing the soft association entirely (SATA-S) drops Squirrel accuracy to 35.23%, below the MLP-level results of the full baselines, showing the attribute-topology coupling is critical.
Limitations
- Reported baseline results are copied from original references rather than re-run under identical settings.
- Accuracy-only evaluation on six datasets; hyperparameter alpha2 strongly affects performance.
Experimental setup
Datasets: Cora Citeseer Photo Computer Squirrel Chameleon
Benchmarks: Cora Citeseer Photo Computer Squirrel Chameleon
Metrics: accuracy
Baselines: GCN GAT GraphSAGE MLP GCNII APPNP JKNet Geom-GCN GPRGNN FAGCN H2GCN GNN-BC
Builds on: GCN APPNP GCNII HSIC-based attribute-topology disentanglement (GNN-BC) Geom-GCN homophily measure
Closest papers by meaning
- Demystifying Structural Disparity in Graph Neural Networks: Can One Size Fit All? (NeurIPS 2023 · similarity 0.90)
- Integrating Co-Training with Edge Discrimination to Enhance Graph Neural Networks Under Heterophily (AAAI 2025 · similarity 0.88)
- Making Classic GNNs Strong Baselines Across Varying Homophily: A Smoothness–Generalization Perspective (NeurIPS 2025 · similarity 0.87)
- What Is Missing For Graph Homophily? Disentangling Graph Homophily For Graph Neural Networks (NeurIPS 2024 · similarity 0.87)
- Improving Graph Neural Networks by Learning Continuous Edge Directions (ICLR 2025 · similarity 0.87)
- GRAIN: Multi-Granular and Implicit Information Aggregation Graph Neural Network for Heterophilous Graphs (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.