Research map / Graph learning / Graph self-supervised learning
Contextual Structure Knowledge Transfer for Graph Neural Networks
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 CS-GNN, a graph transfer learning method that uses ego-networks, moment-based feature smoothness to guide contextual attention, and a group-wise fairness loss to handle homophily shift across domains.
Problem
Graph transfer learning from a labeled source graph to a target graph suffers when homophily structure diverges between domains (homophily shift), degrading GNN performance.
Approach
Builds k-hop ego-networks per node and computes label-free moment-based feature smoothness matrices to produce node-specific contextual attention for message passing, with METIS-based group-wise fairness loss balancing structural patterns.
Key results
- CS-GNN achieves the best average rank (1.0) in direct transfer on Citation, Social, WebKB, Airport, ARXIV and CORA, e.g., 78.39% on A→D Citation vs. 75.44% for UDAGCN.
- CS-GNN outperforms GNN backbones by 5.8% on average in absolute accuracy under direct transfer.
- With fine-tuning at p=0.1% labeled target nodes, CS-GNN gains roughly 10% accuracy over baselines on ACM→DBLP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CS-GNN instead of the strongest reported baseline | Graph transfer learning · Citation, Social, WebKB, Airport, ARXIV, and CORA | accuracy | Helped |
| Add ego-network construction (ablation) | CS-GNN · Citation | accuracy | Helped |
| Add contextual message passing instead of standard message passing (ablation) | CS-GNN · Citation | accuracy | Helped |
| Add group-wise fairness loss instead of standard cross-entropy loss (ablation) | CS-GNN · Citation | accuracy | Helped |
| Increase feature smoothness moments from 1 to 2 (ablation) | CS-GNN · Citation Network | accuracy | Helped |
| Increase feature smoothness moments beyond 2 (ablation) | CS-GNN · Citation Network | accuracy | No clear effect |
| Vary ego-network hop count from 1 to 5 (ablation) | CS-GNN · Citation Network | accuracy | Mixed |
| Vary graph partition number from 2 to 32 (ablation) | CS-GNN · Citation Network | accuracy | Helped |
| Vary temperature from 0.1 to 100 (ablation) | CS-GNN · Citation Network | accuracy | Mixed |
Surprising findings
- Existing graph transfer methods such as DANE, SpecReg and StruRW lose up to 10% accuracy as homophily discrepancy increases.
Limitations
- Hyperparameters (moments, hop, partitions, temperature) require tuning and sensitivity varies.
- Most compared transfer methods require the target graph unavailable during pre-training, which CS-GNN's setting also depends on the design of.
Experimental setup
Datasets: Citation Social WebKB Airport ARXIV CORA ACM DBLP
Benchmarks: Citation Social WebKB Airport ARXIV CORA ACM-DBLP
Metrics: accuracy average rank
Baselines: GCN GAT GCNII JKNet EERM MIXUP DANE UDAGCN GRADE STRURW
Builds on: GCN graph attention networks METIS DANE GRADE
Closest papers by meaning
- One for All: Universal Topological Primitive Transfer for Graph Structure Learning (NeurIPS 2025 · similarity 0.91)
- Demystifying Structural Disparity in Graph Neural Networks: Can One Size Fit All? (NeurIPS 2023 · similarity 0.90)
- GraphBridge: Towards Arbitrary Transfer Learning in GNNs (ICLR 2025 · similarity 0.89)
- How Much Can Transfer? BRIDGE: Bounded Multi-Domain Graph Foundation Model with Generalization Guarantees (ICML 2025 · similarity 0.89)
- The Underappreciated Power of Vision Models for Graph Structural Understanding (NeurIPS 2025 · similarity 0.89)
- What Is Missing For Graph Homophily? Disentangling Graph Homophily For Graph Neural Networks (NeurIPS 2024 · similarity 0.88)
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.