Research map / Label-efficient and robust learning / Domain adaptation and generalization
Improving Domain Generalization with Domain Relations
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 D3G, which trains one head per training domain and reweights these domain-specific predictors at test time using relations derived from domain metadata (refined by a learned similarity), with a relation-aware consistency loss.
Problem
Domain generalization under domain shift where test domains are unseen; prior methods learn a single domain-invariant model, and it is unclear how to build a domain-specific model for a new test domain without its training data.
Approach
Learn one predictive head per training domain and, at test time, predict on a new domain by a relation-weighted average of training-domain heads, where relations come from domain metadata (fixed) combined with learned metadata similarities; a consistency loss weighted by relations regularizes heads for data-scarce domains.
Key results
- D3G achieves MSE 0.342 on TPT-48 N->S versus 0.371 for the best baseline (TRO), and 0.236 on E->W versus 0.262 for IRM.
- On ChEMBL-STRING PPI>50, D3G reaches 78.67% ROC-AUC versus 75.42% for the best baseline (CORAL and SelfReg).
- On FMoW-WILDS, D3G reaches 39.47% worst-region accuracy versus 37.64% for V-REx, and the paper reports an average improvement of 10.6% over the best prior method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use D³G instead of the strongest listed baseline | TPT-48 N (24) → S (24) · TPT-48 N (24) → S (24) | MSE: 0.371 ± 0.054 -> 0.342 ± 0.019 | Helped |
| use D³G instead of the strongest listed baseline | TPT-48 E (24) → W (24) · TPT-48 E (24) → W (24) | MSE: 0.262 $\pm$ 0.034 -> 0.236 $\pm$ 0.063 | Helped |
| use D³G instead of the strongest listed baseline | FMoW-Asia · FMoW-Asia | Worst accuracy: 26.99 $\pm$ 1.27% -> 28.12 $\pm$ 0.28% | Helped |
| use D³G instead of the strongest listed baseline | FMoW-WILDS · FMoW-WILDS | Worst accuracy: 37.64 $\pm$ 0.92% -> 39.47 $ \pm$ 0.57% | Helped |
| use D³G instead of the strongest listed baseline | ChEMBL-STRING PPI>50 · ChEMBL-STRING PPI>50 | ROC-AUC: 75.42 $\pm$ 0.15% -> 78.67 ± 0.16% | Helped |
| use D³G instead of the strongest listed baseline | ChEMBL-STRING PPI>100 · ChEMBL-STRING PPI>100 | ROC-AUC: 73.32 $\pm$ 0.21% -> 77.24 $\pm$ 0.30% | Helped |
| add fixed domain relations to prediction weighting (ablation) | D³G · FMoW-WILDS | Worst accuracy: 35.32 ± 0.66% -> 39.37 ± 0.34% | Helped |
| add learned relations without fixed relations (ablation) | D³G · FMoW-Asia | Worst accuracy: 26.93 ± 0.47% -> 21.18 ± 2.30% | Hurt |
| combine fixed and learned domain relations (ablation) | D³G · ChEMBL-STRING PPI>100 | ROC-AUC: 73.38 ± 0.13% -> 77.24 ± 0.30% | Helped |
| add domain-aware consistency regularization (ablation) | D³G with fixed relations · FMoW and ChEMBL-STRING | performance | Helped |
Surprising findings
- Using learned relations alone performs worse than using no relations on FMoW-Asia (21.18% vs 26.93% worst accuracy), which the authors attribute to the lack of informative signals without fixed relations.
- Many invariant-learning baselines (e.g., IRM, IB-IRM) are inconsistent versus ERM, performing well on some datasets and poorly on others (e.g., IRM at 51.73% ROC-AUC on PPI>100 vs ERM 71.91%).
Limitations
- Theory relies on strong assumptions: uniform metadata representations on [0,1]^r, Lipschitz heads, and a bandwidth-based relation indicator.
- Performance degrades for test domains without nearby training domains, as noted for DG-15.
- Requires domain IDs and informative domain metadata to construct relations; relation learning alone is unreliable.
Experimental setup
Datasets: DG-15 TPT-48 FMoW-Asia FMoW-WILDS ChEMBL-STRING
Benchmarks: DG-15 TPT-48 FMoW-Asia FMoW-WILDS ChEMBL-STRING
Metrics: MSE worst-group accuracy average accuracy ROC-AUC accuracy
Baselines: ERM GroupDRO IRM IB-IRM IB-ERM V-REx DANN CORAL MMD RSC CAD SelfReg
Builds on: empirical risk minimization IRM CORAL DRM LLE
Closest papers by meaning
- Unknown Domain Inconsistency Minimization for Domain Generalization (ICLR 2024 · similarity 0.88)
- Is Large-scale Pretraining the Secret to Good Domain Generalization? (ICLR 2025 · similarity 0.88)
- Federated Domain Generalization with Data-free On-server Matching Gradient (ICLR 2025 · similarity 0.88)
- One-Step Generalization Ratio Guided Optimization for Domain Generalization (ICML 2025 · similarity 0.87)
- Set Valued Predictions For Robust Domain Generalization (ICML 2025 · similarity 0.87)
- Towards Generalizing to Unseen Domains with Few Labels (CVPR 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.