Research map / Graph learning / Graph self-supervised learning
Enhancing Graph Invariant Learning from a Negative Inference Perspective
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 NeGo, a graph OOD framework that infers environments via a class-specific negative prompt (treating extra-class samples as environments) and uses the inferred environment embedding to guide invariant subgraph extraction.
Problem
Graph out-of-distribution generalization under complex environment shifts, where existing environment-centered invariant learning methods treat spurious subgraphs as the environment and thus have a narrow environment inference space.
Approach
Replace positive environment inference (spurious subgraphs as environments) with negative inference: a class-specific learnable negative prompter (cross-attention over a GNN graph embedding) models all extra-class samples as environments, and an attention-based residual interaction injects the inferred environment embedding into node representations before causal subgraph extraction.
Key results
- NeGo achieves 75.21 ROC-AUC (ID val) on GOOD-HIV-scaffold vs 74.04 for LECI, the strongest baseline in that column.
- On GOOD-Twitter, NeGo reaches 60.82 (ID val) and 61.25 (OOD val), versus 59.35 and 59.64 for LECI.
- On the complex-environment SPURIOUS-MOTIF variant, NeGo reaches 87.34% on the original and 80.29% on the modified dataset, with only a minor drop.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NeGo instead of LECI | Graph OOD prediction · GOOD-HIV-scaffold ID val | ROC-AUC: 74.04±0.65 -> 75.21±0.73* | Helped |
| Use NeGo instead of LECI | Graph OOD prediction · GOOD-HIV-size OOD val | ROC-AUC: 65.44±1.78 -> 65.92±1.82* | Helped |
| Use NeGo instead of LECI | Graph OOD prediction · DrugOOD-assay OOD val | ROC-AUC: 73.45±0.17 -> 73.94±0.25* | Helped |
| Use NeGo instead of LECI | Graph OOD prediction · DrugOOD-size ID val | ROC-AUC: 65.93±0.43 -> 65.49±0.73 | Hurt |
| Use NeGo instead of LECI | Graph OOD prediction · GOOD-SST2 ID val | accuracy: 82.93±0.22 -> 82.72±0.51 | Hurt |
| Use NeGo instead of LECI | Graph OOD prediction · GOOD-SST2 OOD val | accuracy: 83.44±0.27 -> 84.16±0.23* | Helped |
| Add the negative prompter and negative loss (ablation) | NeGo · DrugOOD-assay | ROC-AUC: 70.37 -> 73.20 | Helped |
| Add the negative prompter and negative loss (ablation) | NeGo · GOOD-Twitter | accuracy: 58.41 -> 60.82 | Helped |
| Integrate inferred environment information during causal-subgraph extraction (ablation) | NeGo · DrugOOD-assay | ROC-AUC: 71.71 -> 73.20 | Helped |
| Integrate inferred environment information during causal-subgraph extraction (ablation) | NeGo · GOOD-Twitter | accuracy: 59.17 -> 60.82 | Helped |
Limitations
- Theoretical propositions rely on the Gaussian environment mixture assumption borrowed from prior work.
- Efficiency is reported as a modest time overhead relative to LECI and GALA rather than a reduction.
- Hardware and code/data availability are not specified in the main text.
Experimental setup
Datasets: GOOD-Motif GOOD-CMNIST GOOD-HIV DrugOOD LBAP-core-ic50 GOOD-SST2 GOOD-Twitter SPURIOUS-MOTIF
Benchmarks: GOOD-HIV-scaffold GOOD-HIV-size DrugOOD-assay DrugOOD-size GOOD-SST2 GOOD-Twitter GOOD-Motif GOOD-CMNIST SPURIOUS-MOTIF
Metrics: ROC-AUC accuracy training time (s/epoch)
Baselines: ERM IRM V-Rex IB-IRM DIR GSAT CAL CIGA GIL LECI GALA
Builds on: IRM DIR GSAT CIGA LECI GALA
Closest papers by meaning
- Improving Out-of-Distribution Generalization in Graphs via Hierarchical Semantic Environments (CVPR 2024 · similarity 0.90)
- Improving Generalization of Dynamic Graph Learning via Environment Prompt (NeurIPS 2024 · similarity 0.90)
- Dissecting the Failure of Invariant Learning on Graphs (NeurIPS 2024 · similarity 0.89)
- Graph Invariant Learning with Subgraph Co-mixup for Out-of-Distribution Generalization (AAAI 2024 · similarity 0.88)
- Disentangling Invariant Subgraph via Variance Contrastive Estimation under Distribution Shifts (ICML 2025 · similarity 0.88)
- Out-of-Distribution Generalization on Graphs via Progressive Inference (AAAI 2025 · 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.