Research map / Graph learning / GNN architectures
Learning to Reweight for Generalizable Graph Neural Network
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 L2R-GNN, which learns per-graph weights via a nonlinear graph decorrelation over clusters of representation variables plus a first-order bi-level optimization, to improve GNN out-of-distribution generalization on graph classification.
Problem
GNNs degrade under distribution shifts between training and testing graphs because they exploit spurious correlations learned under the i.i.d. assumption, including size shifts and scaffold splits on molecular graphs.
Approach
Instead of decorrelating all pairs of representation variables (as StableGNN and OOD-GNN do), cluster variables by the stability of their correlations and learn sample weights that remove dependencies only between different clusters, with a momentum weight queue and first-order bi-level optimization to avoid overfitting.
Key results
- On six OGB molecular datasets, L2R-GNN reports the best scores in every column, e.g., 78.6 ROC-AUC on TOX21 and 91.9 on CLINTOX, outperforming OOD-GNN and StableGNN.
- On size-split graph classification, L2R-GNN achieves 68.2% on COLLAB, 78.9% on PROTEINS and 80.8% on D&D, the best among listed baselines.
- Ablation shows removing bi-level optimization or the graph decorrelation module lowers accuracy on synthetic biased data, and the first-order approximation performs best.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply L2R-GNN instead of OOD-GNN | Graph property prediction · TOX21 | ROC-AUC: 76.2±1.3 -> 78.6±1.2 | Helped |
| Apply L2R-GNN instead of StableGNN | Graph property prediction · BACE | ROC-AUC: 79.2±2.4 -> 81.9±1.0 | Helped |
| Apply L2R-GNN instead of OOD-GNN | Graph property prediction · BBBP | ROC-AUC: 68.4±2.8 -> 70.6±1.3 | Helped |
| Apply L2R-GNN instead of OOD-GNN | Graph property prediction · CLINTOX | ROC-AUC: 89.1±1.6 -> 91.9±1.5 | Helped |
| Apply L2R-GNN instead of OOD-GNN | Graph property prediction · HIV | ROC-AUC: 78.2±1.2 -> 79.7±1.0 | Helped |
| Apply L2R-GNN instead of OOD-GNN | Graph property prediction · ESOL | RMSE: 0.94±0.06 -> 0.84±0.07 | Helped |
| Apply L2R-GNN instead of StableGNN | Graph classification with graph-size distribution shift · COLLAB | accuracy: 67.3±1.4 -> 68.2±1.4 | Helped |
| Apply L2R-GNN instead of OOD-GNN | Graph classification with graph-size distribution shift · PROTEINS | accuracy: 77.1±1.1 -> 78.9±0.7 | Helped |
| Apply L2R-GNN instead of OOD-GNN | Graph classification with graph-size distribution shift · D&D | accuracy: 79.1±1.3 -> 80.8±1.2 | Helped |
| Add bi-level optimization (ablation) | L2R-GNN · D&D | validation loss | Helped |
| Add graph decorrelation (ablation) | L2R-GNN · Graph classification | accuracy | Helped |
Surprising findings
- Second-order bi-level approximation was not better than first-order in the reported experiments.
- FactorGCN performs much worse than standard GNNs on several benchmarks (e.g., 56.2 on TOX21).
Limitations
- Decorrelation and bi-level optimization add hyperparameters (number of clusters K, queue size, momentum) whose sensitivity drives performance.
- Evaluation is limited to graph classification and regression on specific distribution shifts (size and scaffold splits, synthetic motif bias).
Experimental setup
Datasets: COLLAB PROTEINS D&D OGBG-MOLTOX21 OGBG-MOLBACE OGBG-MOLBBBP OGBG-MOLCLINTOX OGBG-MOLHIV OGBG-MOLESOL synthetic wheel-motif graphs
Benchmarks: COLLAB PROTEINS D&D TOX21 BACE BBBP CLINTOX HIV ESOL synthetic wheel-motif dataset
Metrics: accuracy ROC-AUC RMSE
Baselines: GCN GIN SGC JKNet FactorGCN PNA TopKPool SAGPool OOD-GNN StableGNN
Builds on: OOD-GNN StableGNN random Fourier features gradient-based meta-learning HSIC
Closest papers by meaning
- Topology-Aware Dynamic Reweighting for Distribution Shifts on Graph (ICML 2025 · similarity 0.90)
- A Manifold Perspective on the Statistical Generalization of Graph Neural Networks (ICML 2025 · similarity 0.89)
- Generalization of Graph Neural Networks Is Robust to Model Mismatch (AAAI 2025 · similarity 0.88)
- Pruning Spurious Subgraphs for Graph Out-of-Distribution Generalization (NeurIPS 2025 · similarity 0.88)
- Dynamic Rescaling for Training GNNs (NeurIPS 2024 · similarity 0.88)
- Online GNN Evaluation Under Test-time Graph Distribution Shifts (ICLR 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.