Research map / Probabilistic and scientific modeling / Causal discovery
Partial Transportability 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.
Introduces partial transportability, bounding target-domain risk under selection diagrams via canonical and neural causal models, with a gradient-based Neural-TR optimizer and CRO for worst-case-optimal classifiers.
Problem
Generalization risk of a classifier in an unseen target domain is often non-transportable (not uniquely determined) from source data and causal assumptions, so only bounds can be given.
Approach
Bound target-domain functionals by optimizing over canonical SCMs (discrete exogenous variables) and, for scalability, over neural causal models with parameter sharing across domains dictated by selection diagrams; decompose by c-components so identifiable factors are plugged in directly, and iteratively train classifiers against worst-case NCMs (CRO).
Key results
- Theorem gives tight upper bounds for target risk with canonical models over discrete SCMs under selection-diagram constraints.
- Neural-TR yields worst-case risk estimates for classifiers that match MCMC-based bounds in the smoking and bow-model examples.
- CRO is shown to return a worst-case optimal classifier in the large-sample limit.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use the unstable feature W in h1(C, W) | h1(C, W) · Classifier example, source domains M1 and M2 versus target domain M* | error: 1% and 4% -> 49% | Hurt |
| use the direct causes C in h2(C) | classifier in the example · Classifier example, M1, M2, and M* | error: 20% -> 20% -> 20% | No clear effect |
| use the invariant descendant feature Z in h3(Z) | classifier in the example · Classifier example, M1, M2, and M* | error: 3% -> 5% -> 4% | No clear effect |
| evaluate the three prediction rules with Neural-TR under the stated discrepancy in S | Neural-TR · Smoking and lung cancer example, target domain M* | risk: close to 0.4 | No clear effect |
Surprising findings
- A classifier using only the descendant Z of Y performs nearly optimally across domains despite not being causal or ERM-optimal.
- The ERM classifier using the unstable feature W incurs 50% loss in one target domain.
Experimental setup
Datasets: CMNIST
Benchmarks: CMNIST
Metrics: worst-case risk upper bound target risk
Baselines: empirical risk minimization causal predictor
Builds on: Neural Causal Models Balke and Pearl canonical parameterization selection diagrams s-admissibility
Closest papers by meaning
- Transportable Representations for Domain Generalization (AAAI 2024 · similarity 0.91)
- Probability-Polarized Optimal Transport for Unsupervised Domain Adaptation (AAAI 2024 · similarity 0.85)
- Theoretical Performance Guarantees for Partial Domain Adaptation via Partial Optimal Transport (ICML 2025 · similarity 0.85)
- PDE+: Enhancing Generalization via PDE with Adaptive Distributional Diffusion (AAAI 2024 · similarity 0.85)
- Generative Conditional Distributions by Neural (Entropic) Optimal Transport (ICML 2024 · similarity 0.85)
- Continuous Domain Generalization (NeurIPS 2025 · similarity 0.85)
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.