Covariate-Shift Generalization via Random Sample Weighting

Shifts in the marginal distribution of covariates from training to test phase, named covariate-shifts, often lead unstable prediction performance across agnostic testing data, especially under model misspecification. Recent literature on invariant learning attempts learn an predictor heterogeneous environments. However, learned depends heavily availability and quality provided In this paper, we propose a simple effective non-parametric method for generating environments via Random Sample Weighting (RSW). Given dataset single source environment, randomly generate set covariate-determining sample weights use each weighted simulate environment. We theoretically show that appropriate conditions, such random weighting can produce sufficient heterogeneity be exploited by common invariance constraints find variables stable covariate shifts. Extensive experiments both simulated real-world datasets clearly validate effectiveness our method.