Research map / Graph learning / Graph self-supervised learning
Empowering Graph Invariance Learning with Deep Spurious Infomax
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 EQuAD, which uses infomax-based self-supervised graph encoders to isolate spurious features, quantifies per-sample spurious correlation via classifier logits, and retrains a GNN from scratch with a decorrelation loss for OOD graph classification.
Problem
Graph invariance learning for out-of-distribution generalization relies on assumptions about the correlation strength between spurious features and labels, which often do not hold under real-world shifts.
Approach
Infomax-based self-supervised learning captures spurious features regardless of spurious-label correlation strength; the paper then quantifies each sample's spurious correlation from logits and trains a GNN from scratch with a reweighted decorrelation objective plus ERM loss.
Key results
- On SPMotif, EQuAD reaches 74.61% accuracy at b=0.33 and 69.47% at b=0.90, staying stable across bias levels where baselines degrade.
- On two-piece graphs with (0.80,0.90) and (0.70,0.90) bias, EQuAD achieves 75.81% and 71.95%, while GREA, DIR-style and environment methods fail catastrophically.
- On DrugOOD EC50-Assay, EQuAD reaches 79.36% AUC versus 77.48% for the next best (iMoLD); the paper reports an average improvement of 31.76% over baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply EQuAD instead of the strongest listed baseline | GIN graph classification · SPMotif across b=0.33, 0.40, 0.60, 0.90 | accuracy: 72.91±1.92 -> 74.61±1.23; 67.96±5.27 -> 73.13±1.56; 67.31±6.84 -> 71.93±1.94; 62.07±2.20 -> 69.47±2.06 | Helped |
| apply EQuAD instead of the strongest listed baseline | GIN graph classification · Two-piece graph datasets | accuracy: 83.65±0.44 -> 82.76±0.71; 62.25±3.71 -> 75.81±0.51; 49.65±3.93 -> 71.95±1.41 | Mixed |
| apply EQuAD instead of the strongest listed baseline | GIN graph classification · DrugOOD IC50-Assay, IC50-Scaffold, IC50-Size, EC50-Assay, EC50-Scaffold | AUC: 72.11±0.51 -> 73.26±0.47; 69.14±0.70 -> 69.78±0.41; 67.92±0.43 -> 68.19±0.24; 77.48±1.70 -> 79.36±0.73; 67.79±0.88 -> 68 | Helped |
| apply EQuAD instead of the strongest listed baseline | GIN graph classification · DrugOOD EC50-Size | AUC: 67.09±0.91 -> 66.37±0.64 | Hurt |
| apply EQuAD instead of the strongest listed baseline | GIN graph classification · MolBACE and MolBBBP | AUC: 82.37±2.37 -> 81.83±2.67; 69.84±1.84 -> 71.22±1.47 | Mixed |
| use model-specific adaptive reweighting instead of averaging logits matrices (ablation) | EQuAD · DrugOOD EC50-Assay | AUC: 78.03±2.24 -> 79.36±0.63 | Helped |
| use model-specific adaptive reweighting instead of averaging logits matrices (ablation) | EQuAD · DrugOOD EC50-Scaffold | AUC: 67.45±0.32 -> 68.12±0.48 | Helped |
| use model-specific adaptive reweighting instead of averaging logits matrices (ablation) | EQuAD · DrugOOD EC50-Size | AUC: 66.13±0.38 -> 66.37±0.64 | Helped |
| use multiple prediction logits matrices instead of a single matrix (ablation) | EQuAD · DrugOOD EC50-Assay | AUC: 77.46±0.29 -> 79.36±0.63 | Helped |
| use multiple prediction logits matrices instead of a single matrix (ablation) | EQuAD · DrugOOD EC50-Scaffold | AUC: 67.68±0.17 -> 68.12±0.48 | Helped |
| use multiple prediction logits matrices instead of a single matrix (ablation) | EQuAD · DrugOOD EC50-Size | AUC: 66.02±0.41 -> 66.37±0.64 | Helped |
| train the encoder from scratch rather than use the infomax-pretrained encoder (ablation) | EQuAD · test accuracy | accuracy | Helped |
Surprising findings
- Infomax-based SSL representations alone yield test accuracy below 40% on SPMotif after reweighting, indicating they predominantly encode spurious features.
- Environment inference and augmentation methods (MoleOOD, GIL, GREA, DIR) fail catastrophically when spurious-label correlation is strong, e.g., on two-piece graphs with (0.80,0.90) bias.
Limitations
- Decorrelation objective can face optimization difficulties under strong spurious-label correlation with minibatch gradient descent, motivating an alternative objective.
- Theoretical guarantees rely on assumptions such as symmetric bins and equal-size subgroups, and the paper's analysis is limited to graph data.
- Experiments are restricted to graph classification with a GIN backbone; more expressive encoders are only discussed as future potential.
Experimental setup
Datasets: SPMotif Two-piece graph DrugOOD MolBACE MolBBBP
Benchmarks: SPMotif Two-piece graph DrugOOD MolBACE MolBBBP
Metrics: accuracy AUC central moment distance
Baselines: ERM IRM V-Rex IB-IRM EIIL GREA GSAT GIL DisC CIGA GALA DIR
Builds on: InfoGraph IRM CIGA DisC GIN
Closest papers by meaning
- Dissecting the Failure of Invariant Learning on Graphs (NeurIPS 2024 · similarity 0.89)
- Disentangling Invariant Subgraph via Variance Contrastive Estimation under Distribution Shifts (ICML 2025 · similarity 0.88)
- Subgraph Invariant Learning Towards Large-Scale Graph Node Classification (AAAI 2025 · similarity 0.88)
- Equivariance Everywhere All At Once: A Recipe for Graph Foundation Models (NeurIPS 2025 · similarity 0.88)
- Pruning Spurious Subgraphs for Graph Out-of-Distribution Generalization (NeurIPS 2025 · similarity 0.87)
- Disentangled Graph Self-supervised Learning for Out-of-Distribution Generalization (ICML 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.