Taming Covariate Shift: From Prediction to Uncertainty Quantification

CONG MA – UNIVERSITY OF CHICAGO

ABSTRACT

Covariate shift arises when the distribution of input variables (covariates) differs between training and test phases, while the relationship between inputs and outputs remains unchanged. If not addressed, this can significantly degrade model performance, leading to unreliable predictions and poor generalization.

In this talk, I will present methods to effectively “tame” covariate shift, with a focus on both improving prediction accuracy and providing rigorous uncertainty quantification. The first part will address the challenge of covariate shift in nonparametric regression, where I will characterize the fundamental limits of the problem and introduce simple, yet effective, procedures that achieve these limits.

In the second part, I will turn to the construction of distribution-free prediction intervals under covariate shift. Specifically, I will present a shape-constrained approach to conformal inference, which ensures valid coverage while producing shorter, more informative prediction intervals.

This is based on joint works with Reese Pathak, Martin J. Wainwright, Yu Gui and Rina F. Barber.