The domain generalization problem is about learning predictive models that are robust to distributional shifts when applied to new, unprecedented environments. Existing methods usually require labeled data from multiple training points, which limits their performance when labeled data is scarce. In this work, we study domain adaptation in a causally opposed environment, where the effect causes the observed covariates. Under this structure, environmental disturbances that affect the covariates do not propagate to the outcome, which promotes the sensitivity of the model to these disturbances. Worst of all, estimating these perturbation directions does not require labels, allowing us to use label-free data from multiple locations. We propose two methods that penalize the sensitivity of the model to the variance of the mean and the covariance of the covariates across locations, respectively, and prove that these methods have the best and worst guarantees under certain geographic classifications. Finally, we demonstrate the intelligent performance of our method on a controlled physical system and physical state signal dataset.
- † An apple
- ‡ ETH Zürich
