Research map / Label-efficient and robust learning / OOD and anomaly detection
GeSS: Benchmarking Geometric Deep Learning under Scientific Applications with Distribution Shifts
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.
Introduces GeSS, a benchmark evaluating geometric deep learning models (EGNN, DGCNN, Point Transformer) and 11 learning algorithms under 10 distribution-shift cases across HEP, materials, and biochemistry, at three levels of OOD information.
Problem
Geometric deep learning models used in scientific applications face distribution shifts between source and target domains, but little systematic study exists of how GDL models handle such shifts or how available OOD information (none, unlabeled features, few labels) affects methods.
Approach
Categorizes scientific distribution shifts via a causal data model (covariate, concept, and two conditional-shift subtypes) and evaluates OOD generalization, domain adaptation, and transfer learning methods across three OOD-information levels on new and adapted geometric datasets.
Key results
- On Signal shift z'10->2mu with EGNN, DANN outperforms ERM by 9.4% and all No-Info OOD generalization methods by at least 6.6%.
- On Fidelity shift HSE06, TL with 1000 labels improves MAE over other methods by at least 40%, while gains are limited for HSE06*.
- Benchmark spans 30 settings (10 shift cases x 3 OOD-info levels) with 3 GDL backbones and 11 learning algorithms, reporting mean and std over 3 replicates.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DANN with OOD feature information instead of ERM without OOD information | EGNN on Track-Signal z′10→2μ · Track-Signal z′10→2μ | ACC: ↑ 9.4% | Helped |
| Use DANN with OOD feature information instead of OOD generalization methods without OOD information | EGNN on Track-Signal z′10→2μ · Track-Signal z′10→2μ | ACC: ↑ at least 6.6% | Helped |
| Use domain adaptation with OOD features rather than ERM | GDL models on Track-Pileup · Track-Pileup | ACC | No clear effect |
| Fine-tune on labeled OOD data rather than use non-fine-tuning methods | GDL models on QMOF HSE06 · QMOF HSE06 | MAE: at least 40% | Helped |
| Fine-tune on a few labeled OOD examples | GDL models on QMOF HSE06* · QMOF HSE06* | MAE | No clear effect |
| Use GroupDRO with source groups split by signal type rather than ERM | EGNN on Track-Signal · Track-Signal | ACC | Helped |
| Use GroupDRO with source groups split by point count rather than ERM | EGNN on Track-Pileup · Track-Pileup | ACC | Hurt |
| Increase the number of labeled OOD fine-tuning examples | GDL models under distribution shifts · Multiple shift cases | Test-OOD performance | Mixed |
Surprising findings
- Fine-tuning with very few labeled OOD samples can hurt performance relative to ERM in cases with smaller shifts, likely due to catastrophic forgetting.
- GroupDRO consistently helps under Signal shift but underperforms ERM under Pileup shift, as the subgroup splits poorly reflect the shift.
Limitations
- Data models cover only a subset of possible causal relationships; other shift mechanisms are out of scope.
- Only three GDL backbones, and only distance-based geometric information is used.
- Label shift is not present in the datasets and is not studied.
Experimental setup
Datasets: Track-Pileup Track-Signal QMOF DrugOOD-3D
Benchmarks: Track-Pileup Track-Signal QMOF DrugOOD-3D
Metrics: ACC AUC MAE
Baselines: ERM VREx MixUp GroupDRO DIR LRI DANN DeepCoral
Builds on: DANN DeepCoral DrugOOD DIR LRI
Closest papers by meaning
- Fast geometric learning with symbolic matrices (NeurIPS 2020 · similarity 0.84)
- Bridging Geometric States via Geometric Diffusion Bridge (NeurIPS 2024 · similarity 0.83)
- GeoTransolver: Learning Physics on Irregular Domains using Multi-scale Geometry Aware Physics Attention Transformer (NeurIPS 2026 · similarity 0.83)
- Large Language-Geometry Model: When LLM meets Equivariance (ICML 2025 · similarity 0.82)
- Geometry Awakening: Cross-Geometry Learning Exhibits Superiority over Individual Structures (NeurIPS 2024 · similarity 0.82)
- Geometric Trajectory Diffusion Models (NeurIPS 2024 · similarity 0.82)
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.