Research map / Graph learning / Graph self-supervised learning
Improving Out-of-Distribution Generalization in Graphs via Hierarchical Semantic Environments
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 hierarchical semantic environment generation via stochastic subgraph extraction and hierarchical environment diversification/consistency losses for graph OOD generalization, evaluated mainly on DrugOOD and GIN backbones.
Problem
Graph out-of-distribution generalization fails when training environments are numerous and heterogeneous (e.g., DrugOOD scaffold splits with ~6881 environments), and existing flat environment generation/inference methods do not capture these complex distribution shifts.
Approach
Generate environments hierarchically: extract variant/invariant subgraphs with Gumbel-Softmax edge sampling, re-extract at each hierarchy to produce progressively coarser environments, and train with an intra-hierarchy diversification loss plus inter-hierarchy contrastive consistency losses before doing invariant learning on the top-level environments.
Key results
- On DrugOOD IC50-SCA, the method reaches 70.72 ROC-AUC vs 69.70 for the best baseline (CIGAV2), a gain of about 1.0 points.
- On EC50-SCA, the method reaches 69.73 ROC-AUC vs 67.98 for GREA and 66.05 for CIGAV1; the abstract claims up to 2.83% improvement on EC50 tasks.
- Flat environment baselines (Real, Rand#2, Infer#2) perform comparably to or worse than ERM on IC50-SCA, while the hierarchical approach improves over them.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use hierarchical semantic environments rather than the strongest baseline | Ours on DrugOOD IC50-ASSAY · DrugOOD IC50-ASSAY | ROC-AUC: 74.01±0.11 vs 73.17±0.39 | Helped |
| Use hierarchical semantic environments rather than the strongest baseline | Ours on DrugOOD IC50-SCA · DrugOOD IC50-SCA | ROC-AUC: 70.72±0.30 vs 69.70±0.27 | Helped |
| Use hierarchical semantic environments rather than the strongest baseline | Ours on DrugOOD IC50-SIZE · DrugOOD IC50-SIZE | ROC-AUC: 68.64±0.23 vs 67.78±0.76 | Helped |
| Use hierarchical semantic environments rather than the strongest baseline | Ours on DrugOOD EC50-ASSAY · DrugOOD EC50-ASSAY | ROC-AUC: 80.82±0.21 vs 79.24±1.36 | Helped |
| Use hierarchical semantic environments rather than the strongest baseline | Ours on DrugOOD EC50-SCA · DrugOOD EC50-SCA | ROC-AUC: 69.73±0.21 vs 67.98±1.00 | Helped |
| Use hierarchical semantic environments rather than the strongest baseline | Ours on DrugOOD EC50-SIZE · DrugOOD EC50-SIZE | ROC-AUC: 66.87±0.38 vs 66.01±0.84 | Helped |
| Use inferred hierarchical environments instead of two randomly assigned environments (ablation) | Graph invariant learning · DrugOOD IC50-SCA | ROC-AUC: 68.54±0.64 -> 70.72±0.30 | Helped |
| Use inferred hierarchical environments instead of two randomly assigned environments (ablation) | Graph invariant learning · DrugOOD IC50-SIZE | ROC-AUC: 67.63±0.33 -> 68.64±0.23 | Helped |
| Use inferred hierarchical environments instead of provided real environments (ablation) | Graph invariant learning · DrugOOD IC50-SCA | ROC-AUC: 68.77±0.72 -> 70.72±0.30 | Helped |
| Use inferred hierarchical environments instead of provided real environments (ablation) | Graph invariant learning · DrugOOD IC50-SIZE | ROC-AUC: 67.60±0.32 -> 68.64±0.23 | Helped |
| Use hierarchical rather than flat inferred environments (ablation) | Graph invariant learning · DrugOOD IC50-SCA | ROC-AUC: 69.14±0.80 -> 70.72±0.30 | Helped |
| Use hierarchical rather than flat inferred environments (ablation) | Graph invariant learning · DrugOOD IC50-SIZE | ROC-AUC: 67.55±0.34 -> 68.64±0.23 | Helped |
Surprising findings
- Many existing invariant learning methods (IRM, V-REX, EIIL, GREA, MoleOOD) perform comparably to or worse than ERM on DrugOOD, especially on the high-environment-count scaffold split.
Limitations
- Gains on several DrugOOD splits are modest (around 1 point) relative to standard deviations.
- Evaluation focuses on DrugOOD; generality to other domains is only briefly claimed in the appendix.
Experimental setup
Datasets: DrugOOD CMNIST-75sp Graph-SST5 Graph-Twitter
Benchmarks: DrugOOD CMNIST-75sp Graph-SST5 Graph-Twitter
Metrics: ROC-AUC accuracy
Baselines: ERM IRM V-REX EIIL IB-IRM GREA CIGA MoleOOD GALA LiSA DisC GIL
Builds on: invariant risk minimization EIIL CIGA GALA InfoNCE
Closest papers by meaning
- Graph Invariant Learning with Subgraph Co-mixup for Out-of-Distribution Generalization (AAAI 2024 · similarity 0.90)
- Enhancing Graph Invariant Learning from a Negative Inference Perspective (ICML 2025 · 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 0.88)
- Dissecting the Failure of Invariant Learning on Graphs (NeurIPS 2024 · similarity 0.88)
- Graph Structure Extrapolation 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.