Research map / Graph learning / Graph self-supervised learning
Disentangled Graph Self-supervised Learning 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 OOD-GCL, a disentangled multi-channel graph encoder pretrained with invariance-regularized contrastive learning (cluster-inferred environments, V-REx) so graph representations generalize under distribution shift without labels.
Problem
Graph OOD generalization typically needs task-dependent labels, which are scarce or unavailable; the paper studies self-supervised graph OOD generalization where invariant and variant factors are entangled in graph representations.
Approach
A disentangled multi-channel graph encoder where each channel's representation is clustered (silhouette-selected k-means) into pseudo-environments, and a per-factor contrastive loss on the other channels is combined with a V-REx invariance penalty so each channel excludes its own latent factor; a linear predictor is then finetuned on the representations.
Key results
- OOD-GCL achieves 76.31% on Motif-base, versus 73.64% for the best baseline (AIA).
- OOD-GCL reaches 60.12% on CMNIST-color, versus 55.85% for AIA.
- On Molhiv scaffold split, OOD-GCL reports 72.86 ROC-AUC, above the best baseline values (e.g., 70.79 for AIA).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OOD-GCL instead of AIA | Graph classification · Motif base | accuracy: 73.64±5.15 -> 76.31±2.37 | Helped |
| Use OOD-GCL instead of AIA | Graph classification · Motif size | accuracy: 55.85±7.98 -> 60.12±2.35 | Helped |
| Use OOD-GCL instead of AIA | Graph classification · CMNIST color | accuracy: 36.37±4.44 -> 41.89±2.03 | Helped |
| Use OOD-GCL instead of AIA | Graph classification · Molbbbp scaffold | ROC-AUC: 70.79±1.53 -> 72.86±1.97 | Helped |
| Use OOD-GCL instead of AIA | Graph classification · Molbbbp size | ROC-AUC: 81.03±5.15 -> 83.12±3.21 | Helped |
| Use OOD-GCL instead of AIA | Graph classification · Molhiv scaffold | ROC-AUC: 71.15±1.81 -> 74.13±2.01 | Helped |
| Use OOD-GCL instead of AIA | Graph classification · Molhiv size | ROC-AUC: 61.64±3.37 -> 63.41±2.59 | Helped |
| Add invariance regularization (ablation) | OOD-GCL · Motif | performance: 72.18 -> 76.31 | Helped |
| Add invariance regularization (ablation) | OOD-GCL · CMNIST | performance: 35.16 -> 41.89 | Helped |
| Add the disentangled module (ablation) | OOD-GCL · Motif | performance: 71.34 -> 76.31 | Helped |
| Add the disentangled module (ablation) | OOD-GCL · CMNIST | performance: 34.98 -> 41.89 | Helped |
Surprising findings
- Graph self-supervised baselines (e.g., InfoGraph, GraphCL) perform comparably to some graph OOD methods on several splits.
Limitations
- Relies on k-means clustering of channel representations as proxy environments, with no ground-truth environment labels.
- Theoretical guarantees rest on mild assumptions and the disentanglement definition; empirical gains are reported on a limited set of four datasets.
- Hyperparameter sensitivity analysis is reported only on CMNIST.
Experimental setup
Datasets: Motif (GOOD) CMNIST (GOOD) Molbbbp (OGB) Molhiv (OGB)
Benchmarks: Motif CMNIST Molbbbp Molhiv
Metrics: accuracy ROC-AUC
Baselines: ERM IRM GroupDRO V-REx DIR CAL GSAT CIGA AIA InfoGraph GraphCL GMI
Builds on: GraphCL V-REx DisenGCN-style multi-channel disentangled GNN (Li et al., 2021a) Wang et al. 2021a disentangled self-supervised learning GOOD benchmark
Closest papers by meaning
- Disentangling Invariant Subgraph via Variance Contrastive Estimation under Distribution Shifts (ICML 2025 · similarity 0.92)
- Pruning Spurious Subgraphs for Graph Out-of-Distribution Generalization (NeurIPS 2025 · similarity 0.90)
- Subgraph Aggregation for Out-of-Distribution Generalization on Graphs (AAAI 2025 · similarity 0.89)
- Enhancing Size Generalization in Graph Neural Networks through Disentangled Representation Learning (ICML 2024 · similarity 0.89)
- Graph Invariant Learning with Subgraph Co-mixup for Out-of-Distribution Generalization (AAAI 2024 · similarity 0.89)
- Improving Out-of-Distribution Generalization in Graphs via Hierarchical Semantic Environments (CVPR 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.