Research map / Label-efficient and robust learning / Domain adaptation and generalization
Do causal predictors generalize better to new domains?
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.
Empirically tests whether predictors using causal features generalize better across domains on 16 tabular prediction tasks, finding that all-feature predictors outperform causal-feature predictors in-domain and out-of-domain.
Problem
Whether models trained on causal features generalize better across domains than models trained on all available features, in realistic tabular datasets with natural domain shifts.
Approach
Shifts the focus from causal algorithms to causal feature sets: defines causal and arguably causal feature selections per task using domain knowledge, then compares standard, domain-generalization and causal ML methods trained on each selection against all features across natural domain shifts.
Key results
- Models using all features achieve the highest in-domain and out-of-domain accuracy, up to error bars, in 16/16 tasks.
- Arguably causal features Pareto-dominate the conservative causal feature set in 16/16 tasks.
- Adding anti-causal features improves out-of-domain accuracy in 5/5 tasks with anti-causal features.
- Causal methods (IRM, REx, IB-IRM, CausIRL, AND-Mask) do not improve upon arguably causal features trained with standard models (8/8 tasks).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use all available features instead of causal features | standard predictors · 16 tabular classification tasks | in-domain accuracy: 16/16 tasks | Helped |
| use all available features instead of causal features | standard predictors · 16 tabular classification tasks | out-of-domain accuracy: 16/16 tasks | Helped |
| expand the conservative causal feature set to arguably causal features | standard predictors · 16 tabular classification tasks | in-domain accuracy: 16/16 tasks | Helped |
| expand the conservative causal feature set to arguably causal features | standard predictors · 16 tabular classification tasks | out-of-domain accuracy: 16/16 tasks | Helped |
| use all available features instead of causal features | standard predictors · 16 tabular classification tasks | shift gap: 7/16 tasks | Helped |
| add anti-causal features to the (arguably) causal feature set | standard predictors · Five tasks with anti-causal features | out-of-domain accuracy: 5/5 tasks | Helped |
| apply causal domain-generalization methods rather than standard models on arguably causal features | IRM, REx, IB-IRM, CausIRL, and AND-Mask · Eight tasks with multiple training domains | out-of-domain accuracy: 8/8 tasks | No clear effect |
| use causal-discovery-selected parents rather than domain-knowledge arguably causal features | ICP, PC, and FCI · Tasks where causal discovery returned parent estimates | out-of-domain accuracy: ICP:~1/1, PC:~4/4, FCI:~1/1 tasks | No clear effect |
| remove one causal feature at a time (ablation) | causal feature subsets · 16 tabular classification tasks | out-of-domain accuracy: one exception in the task `ASSISTments' | No clear effect |
| remove the skill feature (ablation) | ASSISTments predictors · ASSISTments | out-of-domain accuracy | Helped |
Surprising findings
- Removing the skill feature in the ASSISTments task improves out-of-domain accuracy, the one exception found, despite skill being a clearly causal variable.
- Causal discovery algorithms (PC, ICP, FCI) rarely output any causal parents, and when they do, their selections are Pareto-dominated by arguably causal features.
Limitations
- Causal feature labels are expert-based approximations and may be wrong or confounded; robustness checks address single-feature misclassification.
- Causal and domain-generalization methods require at least two training domains, so only eight tasks could test them.
- Findings are limited to tabular datasets; generalization to other modalities such as images is untested.
Experimental setup
Datasets: TableShift (ACS, BRFSS, MIMIC-III, UCI, NHANES, PhysioNet, ANES, ED, Kaggle ASSISTments) MEPS SIPP
Benchmarks: TableShift MEPS SIPP
Metrics: accuracy balanced accuracy shift gap
Baselines: XGBoost LightGBM SAINT TabTransformer NODE FT Transformer IRM REx IB-IRM CausIRL AND-Mask ICP
Builds on: TableShift Invariant Risk Minimization Risk Extrapolation invariant causal prediction PC algorithm
Closest papers by meaning
- Generalization Bounds for Causal Regression: Insights, Guarantees and Sensitivity Analysis (ICML 2024 · similarity 0.86)
- Prediction-powered Generalization of Causal Inferences (ICML 2024 · similarity 0.86)
- Learning Time-Aware Causal Representation for Model Generalization in Evolving Domains (ICML 2025 · similarity 0.86)
- Towards Causal Foundation Model: on Duality between Optimal Balancing and Attention (ICML 2024 · similarity 0.85)
- Can Large Language Models Infer Causation from Correlation? (ICLR 2024 · similarity 0.85)
- Causality can systematically address the monsters under the bench(marks) (NeurIPS 2026 · 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.