Research map / Graph learning / Graph self-supervised learning
Graph Invariant Learning with Subgraph Co-mixup for Out-of-Distribution Generalization
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 IGM, a graph invariant learning method that jointly applies environment Mixup (mixing variant environment subgraphs with invariant subgraphs of other labels) and invariant Mixup (manifold mixup on invariant subgraphs) for graph OOD generalization.
Problem
Graph invariant learning for OOD generalization relies on predefined or inaccurately generated environment partitions, which are hard to obtain, yielding suboptimal OOD performance under size, degree, and structure shifts.
Approach
Learn a subgraph extractor that splits each graph into invariant and environment subgraphs, then generate diverse environments by recombining invariant subgraphs with environment subgraphs of different labels (environment Mixup), and apply manifold mixup only on invariant subgraphs (invariant Mixup), with the two mixups mutually reinforcing each other.
Key results
- On SPMotif, IGM reaches 82.36% accuracy at bias 0.33 versus 77.33% for the next best (CIGA), per the reported table.
- On real-world datasets, IGM reports an average 7.7% improvement over SOTA; the best results on all eight shown datasets (e.g., Graph-SST5 46.69% ACC vs 44.71% for CIGA).
- Ablation shows the full co-mixup outperforms the pre-trained-extractor variant by 4.6% on average and each single mixup still beats ERM by 1.6% on average.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use IGM instead of the strongest baseline | Graph OOD classification · SPMotif (bias 0.33, 0.6, and 0.9) | accuracy: 13.1% average improvement over CIGA | Helped |
| Use IGM instead of the strongest baseline | Graph OOD classification · Real-world datasets | OOD performance: 7.7% | Helped |
| Add both environment Mixup and invariant Mixup (ablation) | IGM with a pre-trained CIGA or DIR subgraph extractor · Graph-Twitter, DD, and BBBP | performance: 4.6% | Helped |
| Add environment Mixup to invariant Mixup alone (ablation) | IGM · Graph-Twitter, DD, and BBBP | performance: not reported | Helped |
| Add invariant Mixup to environment Mixup alone (ablation) | IGM · Graph-Twitter, DD, and BBBP | performance: not reported | Helped |
| Use only invariant Mixup rather than both Mixup modules (ablation) | IGM · Graph-Twitter, DD, and BBBP | performance: not reported | Hurt |
| Use only environment Mixup rather than both Mixup modules (ablation) | IGM · Graph-Twitter, DD, and BBBB | performance: not reported | Hurt |
Surprising findings
- Standard ERM and some invariant learning baselines (e.g., IRM, DIR) frequently underperform ERM on real-world OOD datasets, and CIGA underperforms ERM on DrugOOD-Scaffold.
Limitations
- Evaluated only on graph classification; node classification and dynamic graphs left to future work.
- Mixing and environment construction rely on a learned subgraph extractor whose quality may limit environment diversity.
- Hyperparameter choices (gamma, mu, delta, r_add, K) add tuning burden.
Experimental setup
Datasets: SPMotif Graph-SST5 Graph-Twitter PROTEINS DD DrugOOD OGB BACE BBBP NCI109
Benchmarks: SPMotif Graph-SST5 Graph-Twitter PROTEINS DD DrugOOD OGB BACE BBBP
Metrics: accuracy MCC ROC-AUC NMI
Baselines: ERM G-Mixup Manifold-Mixup IRM V-REx EIIL DIR GSAT CIGA
Builds on: IRM V-REx IRMX Manifold Mixup DIR CIGA GSAT
Closest papers by meaning
- Disentangling Invariant Subgraph via Variance Contrastive Estimation under Distribution Shifts (ICML 2025 · similarity 0.91)
- Improving Out-of-Distribution Generalization in Graphs via Hierarchical Semantic Environments (CVPR 2024 · similarity 0.90)
- Dissecting the Failure of Invariant Learning on Graphs (NeurIPS 2024 · similarity 0.90)
- Subgraph Aggregation for Out-of-Distribution Generalization on Graphs (AAAI 2025 · similarity 0.89)
- Disentangled Graph Self-supervised Learning for Out-of-Distribution Generalization (ICML 2024 · similarity 0.89)
- AGMixup: Adaptive Graph Mixup for Semi-supervised Node Classification (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.