Research map / Federated learning and privacy / Personalized and heterogeneous FL
Federated Self-Explaining GNNs with Anti-shortcut Augmentations
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 FedGR, a federated graph rationalization method that uses complement-aware and difference-aware augmenters to generate shortcut-conflicted samples and learn self-explaining GNNs without shortcuts.
Problem
Graph rationalization methods for explainable GNN graph classification rely on shortcuts in data; existing de-shortcut methods are designed for centralized data and are under-explored in federated learning, where clients have client-specific shortcuts.
Approach
Two augmenters produce shortcut-conflicted samples per client: a complement-aware augmenter that merges a rationale with randomly permuted complements (enforced by a contrastive constraint for sufficiency and independence), and a difference-aware augmenter that masks node features so samples cross the local model's decision boundary while keeping the global model's prediction.
Key results
- On Spurious-Motif with GIN, FedGR reaches 0.4610 / 0.5538 / 0.4977 ACC at bias 0.5 / 0.7 / 0.9, above the best baseline DARE (0.4483 / 0.4891 / 0.4288) at bias 0.5 and 0.9.
- On MolSIDER (GIN), FedGR achieves AUC 0.6017, versus 0.5952 for GSAT and 0.5906 for RGDA.
- Plugging the difference-aware augmenter into DisC, RGDA, GSAT, InterRAT, and DARE improves AUC on all four OGB datasets, by 0.27% to 1.67%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the difference-aware augmenter (ablation) | FedGR · Spurious-Motif bias=0.5 | ACC: 0.4493 ± 0.0238 -> 0.4610 ± 0.0289 | Helped |
| add the difference-aware augmenter (ablation) | FedGR · Spurious-Motif bias=0.7 | ACC: 0.5293 ± 0.0483 -> 0.5538 ± 0.0398 | Helped |
| add the difference-aware augmenter (ablation) | FedGR · Spurious-Motif bias=0.9 | ACC: 0.4333 ± 0.0471 -> 0.4977 ± 0.0315 | Helped |
| add the difference-aware augmenter (ablation) | FedGR · MolHIV | AUC: 0.7214 ± 0.0124 -> 0.7387 ± 0.0186 | Helped |
| add the difference-aware augmenter (ablation) | FedGR · MolToxCast | AUC: 0.6222 ± 0.0055 -> 0.6316 ± 0.0054 | Helped |
| add the difference-aware augmenter (ablation) | FedGR · MolBBBP | AUC: 0.6623 ± 0.0033 -> 0.6690 ± 0.0174 | Helped |
| add the difference-aware augmenter (ablation) | FedGR · MolSIDER | AUC: 0.5886 ± 0.0047 -> 0.6017 ± 0.0202 | Helped |
| add the complement-aware augmenter (ablation) | FedGR · Spurious-Motif bias=0.5 | ACC: 0.4571 ± 0.0372 -> 0.4610 ± 0.0289 | Helped |
| add the complement-aware augmenter (ablation) | FedGR · Spurious-Motif bias=0.7 | ACC: 0.5438 ± 0.0551 -> 0.5538 ± 0.0398 | Helped |
| add the complement-aware augmenter (ablation) | FedGR · Spurious-Motif bias=0.9 | ACC: 0.4682 ± 0.0388 -> 0.4977 ± 0.0315 | Helped |
| add the complement-aware augmenter (ablation) | FedGR · MolHIV | AUC: 0.7321 ± 0.0233 -> 0.7387 ± 0.0186 | Helped |
| add the complement-aware augmenter (ablation) | FedGR · MolToxCast | AUC: 0.6298 ± 0.0035 -> 0.6316 ± 0.0054 | Helped |
Surprising findings
- Several existing de-shortcut methods (DIR, CAL, InterRAT) perform similarly to a plain GIN backbone when transferred to federated settings.
Limitations
- The difference-aware augmenter only perturbs node features and does not alter graph edges, to avoid complexity.
- Assumes the global model is less biased than local models (Assumption 3.3), which is not verified in general.
Experimental setup
Datasets: Spurious-Motif OGB MolHIV OGB MolToxCast OGB MolBBBP OGB MolSIDER
Benchmarks: Spurious-Motif OGB MolHIV OGB MolToxCast OGB MolBBBP OGB MolSIDER
Metrics: ACC AUC Precision@5
Baselines: Vanilla GR DIR DisC CAL GSAT DARE InterRAT RGDA
Builds on: Vanilla graph rationalization DIR CAL GSAT FedAvg
Closest papers by meaning
- Towards Effective Federated Graph Foundation Model via Mitigating Knowledge Entanglement (NeurIPS 2025 · similarity 0.87)
- Self-Interpretable Graph Learning with Sufficient and Necessary Explanations (AAAI 2024 · similarity 0.86)
- Beyond Topological Self-Explainable GNNs: A Formal Explainability Perspective (ICML 2025 · similarity 0.86)
- Factorized Explainer for Graph Neural Networks (AAAI 2024 · similarity 0.86)
- FedSPA: Generalizable Federated Graph Learning under Homophily Heterogeneity (CVPR 2025 · similarity 0.86)
- Less is More: Federated Graph Learning with Alleviating Topology Heterogeneity from A Causal Perspective (ICML 2025 · similarity 0.86)
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.