Research map / Label-efficient and robust learning / Domain adaptation and generalization
Transportable Representations for Domain Generalization
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.
Frames domain generalization as causal transportability of representations, giving graphical and data-driven criteria (invariance property) for when a representation's label score can be transported from source to target domains.
Problem
Domain generalization requires predicting labels in an unseen target domain from labeled source domains; it is unclear which representations of the covariates allow the conditional label distribution to be computed for the target.
Approach
Defines transportable representations: a representation R=φ(X) is transportable if E_target[Y|R=r] is uniquely computable from source distributions given selection diagrams. Extends gTR to queries over representations, and shows that under stability of mechanisms and r-faithfulness, transportability is equivalent to the source invariance property (sound and complete), so invariant representations
Key results
- In synthetic experiments, the transportability-based classifier outperforms INV and ERM, with the gap growing for larger sample sizes (Figure 3).
- ERM performs only slightly better than random guessing and does not improve with more data due to source-target mismatch.
- Theorem 2 shows that under r-faithfulness and stability of mechanisms, a representation's score is transportable iff it satisfies the source invariance property.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the transportability-based classifier instead of ERM or INV | synthetic domain generalization experiment · synthetic target domain | cross-entropy | Helped |
| use INV (regress on the best invariant representation) instead of ERM | synthetic domain generalization experiment · synthetic target domain | cross-entropy | Helped |
| use ERM instead of a random guess | synthetic domain generalization experiment · synthetic target domain | cross-entropy | Helped |
| increase the sample size while using ERM | ERM · synthetic target domain | cross-entropy | No clear effect |
Surprising findings
- Causal features are not always preferable: a transportable representation (Eq. 21) can use non-causal covariates and have higher predictive power than the causal covariate X1 alone.
Limitations
- Requires expert-provided selection diagrams for the graphical approach, which may be hard to obtain.
- Data-driven SoM assumption is conservative and cannot reject source domains as possible targets.
- Assumes infinite samples; finite-sample estimation is outside scope and experiments are synthetic only.
Experimental setup
Datasets: synthetic data from randomly generated SCMs
Benchmarks: synthetic SCM experiment (Example 4 setup)
Metrics: cross-entropy under target distribution
Baselines: ERM INV (invariant representation regression)
Builds on: transportability theory (Bareinboim and Pearl) gTR algorithm (Lee, Correa, Bareinboim 2020) c-faithfulness (Jaber et al. 2020) invariant risk minimization (Arjovsky et al. 2019) Wald et al. 2021 multi-calibration
Closest papers by meaning
- Partial Transportability for Domain Generalization (NeurIPS 2024 · similarity 0.91)
- Causality Inspired Representation Learning for Domain Generalization (CVPR 2022 · similarity 0.88)
- Learning Linear Causal Representations from General Environments: Identifiability and Intrinsic Ambiguity (NeurIPS 2024 · similarity 0.84)
- Causal Representation Learning from Multiple Distributions: A General Setting (ICML 2024 · similarity 0.84)
- Identifiable Exchangeable Mechanisms for Causal Structure and Representation Learning (ICLR 2025 · similarity 0.84)
- Probability-Polarized Optimal Transport for Unsupervised Domain Adaptation (AAAI 2024 · similarity 0.84)
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.