Distributionally Robust Linear Regression With Block Lewis Weights
Organizations: TTIC · Institute for Foundations of Data Science, Yale University
Abstract
We present an algorithm for the group distributionally robust (GDR) least squares problem. Given groups, a parameter vector in , and stacked design matrices and responses and , our algorithm obtains a -multiplicative optimal solution using linear-system-solves of matrices of the form for block-diagonal . Our technical methods follow from a recent geometric construction, block Lewis weights, that relates the empirical GDR problem to a carefully chosen least squares problem and an application of accelerated proximal methods. Our algorithm improves over known interior point methods for moderate accuracy regimes and matches the state-of-the-art guarantees for the special case of regression. We also give algorithms that smoothly interpolate between minimizing the average least squares loss and the distributionally robust loss.