Domain generalization problems are concerned with learning predictive models that are robust to distributional changes when introduced into new, previously unseen environments. Existing methods typically require labeled data from multiple training environments, which limits their applicability when labeled data is lacking. In this study, we study domain generalization in a countercausal setting, where outcomes cause observed covariates. Under this structure, perturbations in the environment that affect the covariates do not propagate to the outcome, thus motivating us to normalize the model’s sensitivity to these perturbations. Importantly, these perturbation direction estimates do not require labels, allowing us to leverage unlabeled data from multiple environments. We each propose two methods that penalize the sensitivity of a model to variations in covariate means and covariances across environments, and prove that these methods guarantee worst-case optimality under certain classes of environments. Finally, we demonstrate the experimental performance of our approach on controlled physical systems and physiological signal datasets.
- † apple
- ‡ ETH Zurich
