Research map / Graph learning / Graph self-supervised learning
Graph Structure Extrapolation 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 G-Splice and FeatX, environment-aware non-Euclidean linear extrapolation data augmentations that generate OOD graphs by splicing graph structures and extrapolating non-causal node features, for graph OOD generalization.
Problem
Graph OOD generalization under structure shifts (size, scaffold/base) and feature shifts; existing graph augmentations mostly interpolate or only produce subgraphs, so they cannot reach OOD regions.
Approach
Defines graph addition (splicing) and subtraction (largest common subgraph removal) to form structural linear extrapolation, uses environment- and label-aware pair learning to extract causal/environmental subgraphs, and generates connecting bridges with a cVAE; feature extrapolation perturbs selected non-causal node features with modulo wrapping into the training domain.
Key results
- G-Splice reaches 86.53% OOD accuracy on GOOD-Motif size, versus 51.74% for ERM, and is best on that split.
- G-Splice+R achieves 65.56 ROC-AUC on GOOD-HIV size and 0.40 MCC on NCI1, the best results among compared methods.
- On FSMotif-S-C, G-Splice+FeatX reaches 75.11% vs. 35.00% for ERM, with the combination outperforming either component alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply G-Splice instead of the strongest listed baseline | graph classification models · GOOD-HIV size | ROC-AUC: 62.38\textpm2.65 -> 64.46\textpm1.38 | Helped |
| apply G-Splice instead of the strongest listed baseline | graph classification models · GOOD-HIV scaffold | ROC-AUC: 71.94\textpm1.31 -> 72.82\textpm1.16 | Helped |
| apply G-Splice instead of the strongest listed baseline | graph classification models · GOOD-SST2 length | accuracy: 82.00\textpm0.73 -> 82.31\textpm0.59 | Helped |
| apply G-Splice instead of the strongest listed baseline | graph classification models · Twitter length | accuracy: 57.67\textpm1.11 -> 58.02\textpm0.40 | Helped |
| apply G-Splice instead of the strongest listed baseline | graph classification models · GOOD-Motif size | accuracy: 54.14\textpm3.11 -> 86.53\textpm2.66 | Helped |
| apply G-Splice instead of the strongest listed baseline | graph classification models · GOOD-Motif base | accuracy: 75.58\textpm0.75 -> 79.86\textpm13.00 | Helped |
| apply G-Splice instead of the strongest listed baseline | graph classification models · DD size | MCC: 0.42\textpm0.03 -> 0.45\textpm0.09 | Helped |
| apply G-Splice instead of the strongest listed baseline | graph classification models · NCI1 size | MCC: 0.30\textpm0.05 -> 0.37\textpm0.08 | Helped |
| add invariant regularization to G-Splice (ablation) | G-Splice · GOOD-HIV size | ROC-AUC: 64.46\textpm1.38 -> 65.56\textpm0.34 | Helped |
| add invariant regularization to G-Splice (ablation) | G-Splice · GOOD-Motif size | accuracy: 86.53\textpm2.66 -> 85.07\textpm4.50 | Hurt |
| add invariant regularization to G-Splice (ablation) | G-Splice · GOOD-Motif base | accuracy: 79.86\textpm13.00 -> 83.96\textpm7.38 | Helped |
| combine FeatX with G-Splice (ablation) | G-Splice · FSMotif size-color | accuracy: 72.11\textpm6.18 -> 75.11\textpm 2.99 | Helped |
Surprising findings
- G-Splice does not always achieve the best in-distribution performance, though it leads on OOD test splits.
- Several strong graph augmentation baselines (e.g., G-Mixup, Graph Mixup) underperform ERM on some structure-shift splits such as GOOD-Motif base and DD.
Limitations
- Theoretical guarantees rest on the causal additivity assumption, which may not hold for all graph classification tasks.
- Combining structure and feature augmentation can introduce extra shifts when the dataset contains only one shift type.
- Hyperparameter selection over augmentation options requires substantial pre-computation and compute resources; link prediction is not covered.
Experimental setup
Datasets: GOOD-HIV GOOD-SST2 Twitter GOOD-Motif DD NCI1 FSMotif
Benchmarks: GOOD-HIV GOOD-SST2 Twitter GOOD-Motif DD NCI1 FSMotif
Metrics: ROC-AUC Matthews correlation coefficient accuracy
Baselines: ERM IRM VREx GroupDRO DANN DIR CIGA GIL LiSA DropNode DropEdge Graph Mixup
Builds on: Mixup invariant risk minimization DIR CIGA LiSA
Closest papers by meaning
- Pruning Spurious Subgraphs for Graph Out-of-Distribution Generalization (NeurIPS 2025 · similarity 0.89)
- Bridging OOD Detection and Generalization: A Graph-Theoretic View (NeurIPS 2024 · similarity 0.89)
- Subgraph Aggregation for Out-of-Distribution Generalization on Graphs (AAAI 2025 · similarity 0.88)
- GOLD: Graph Out-of-Distribution Detection via Implicit Adversarial Latent Generation (ICLR 2025 · similarity 0.87)
- Revisiting Score Propagation in Graph Out-of-Distribution Detection (NeurIPS 2024 · similarity 0.87)
- Improving Out-of-Distribution Generalization in Graphs via Hierarchical Semantic Environments (CVPR 2024 · similarity 0.87)
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.