Research map / Graph learning / Graph self-supervised learning
Out-of-Distribution Generalization on Graphs via Progressive Inference
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 GPro, a graph OOD generalization model that stacks progressive attention-based substructure inference steps with counterfactual feature augmentation and a supervised contrastive causal loss, evaluated on biased graph classification benchmarks.
Problem
Out-of-distribution generalization of GNNs for graph classification under spurious correlations: single-step causal substructure extraction leaves a large gap between learned and ground-truth causal features.
Approach
Decompose causal invariant substructure learning into multiple progressive inference steps, each peeling off high-confidence non-causal edges via an attention-based mask, run with a dual causal/non-causal tower; enlarge training distribution with counterfactual non-causal representations (permutation and mean/variance swapping) and add a graph-level supervised contrastive loss.
Key results
- GPro outperforms the best OOD baseline by 4.91% on average across the three biased graph classification datasets, and up to 6.86% on more severe shifts.
- On CMNIST-75sp at bias 0.95, GPro reaches 79.34% accuracy versus 69.18% for CAL with GAT.
- Ablation: removing the counterfactual causal loss L_cou drops accuracy by about 7.15% on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GPro instead of the strongest listed baseline | graph classification · CMNIST-75sp, CFashion-75sp, and CKuzushiji-75sp; bias 0.8, 0.9, and 0.95 | accuracy (%): GPro outperforms state-of-the-art methods by 4.91% on average | Helped |
| Increase bias severity in the training data | OOD baseline methods · the three graph datasets | accuracy (%) | Hurt |
| Add non-causal representation permutation (ablation) | GPro · CMNIST-75sp, CFashion-75sp, and CKuzushiji-75sp, bias 0.9 | accuracy (%): ↓0.74, ↓1.22, ↓0.72 | Helped |
| Add non-causal representation mean-variance swapping (ablation) | GPro · CMNIST-75sp, CFashion-75sp, and CKuzushiji-75sp, bias 0.9 | accuracy (%): ↓0.87, ↓1.61, ↓1.62 | Helped |
| Add counterfactual loss L_cou (ablation) | GPro · CMNIST-75sp, CFashion-75sp, and CKuzushiji-75sp, bias 0.9 | accuracy (%): ↓4.59, ↓6.75, ↓10.10 | Helped |
| Add supervised contrastive loss L_scl (ablation) | GPro · CMNIST-75sp, CFashion-75sp, and CKuzushiji-75sp, bias 0.9 | accuracy (%): ↓1.33, ↓1.54, ↓0.62 | Helped |
| Add consistency loss L_con (ablation) | GPro · CMNIST-75sp, CFashion-75sp, and CKuzushiji-75sp, bias 0.9 | accuracy (%): ↓0.91, ↓0.44, ↓0.78 | Helped |
| Use more progressive inference steps | GPro · CMNIST-75sp-0.95, CFashion-75sp-0.95, CKuzushiji-75sp-0.95 | accuracy (%): 4.02% improvement with a single inference step over the leading model | Helped |
| Use GraphSAGE as the GPro encoder | GPro · CMNIST-75sp, CFashion-75sp, and CKuzushiji-75sp, bias 0.9 | accuracy (%): 4.03%, 4.35%, and 3.57% compared to the best baseline | Mixed |
| Increase λ3 above 1 | GPro · CMNIST-75sp-0.9, CFashion-75sp-0.9, and CKuzushiji-75sp-0.9 | accuracy (%) | Hurt |
| Set contrastive-loss temperature to 0.5 or 0.9 | GPro · CMNIST-75sp-0.9, CFashion-75sp-0.9, and CKuzushiji-75sp-0.9 | accuracy (%) | Hurt |
Limitations
- Training time is about 1.5x that of CIGA and GALA due to the multi-step inference.
- Evaluated only on three synthetic colored-MNIST-style superpixel datasets plus DrugOOD, so real-world coverage is limited.
- Several hyperparameters (lambda1-3, tau, q, rho) must be tuned.
Experimental setup
Datasets: CMNIST-75sp CFashion-75sp CKuzushiji-75sp DrugOOD
Benchmarks: CMNIST-75sp CFashion-75sp CKuzushiji-75sp DrugOOD
Metrics: accuracy
Baselines: GCN GIN GCNII FactorGCN DiffPool LDD StableGNN CAL DisC CIGA GALA
Builds on: CAL DisC CIGA GCN
Closest papers by meaning
- Quantifying Distributional Invariance in Causal Subgraph for IRM-Free Graph Generalization (NeurIPS 2025 · similarity 0.93)
- A Recipe for Causal Graph Regression: Confounding Effects Revisited (ICML 2025 · similarity 0.89)
- Rethinking Causal Relationships Learning in Graph Neural Networks (AAAI 2024 · similarity 0.89)
- Dissecting the Failure of Invariant Learning on Graphs (NeurIPS 2024 · similarity 0.89)
- Graph Contrastive Invariant Learning from the Causal Perspective (AAAI 2024 · similarity 0.88)
- Subgraph Aggregation for Out-of-Distribution Generalization on Graphs (AAAI 2025 · 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.