Abstract
In multi-source learning with discrete labels, distributional heterogeneityacross domains poses a central challenge to developing predictive models thattransfer reliably to unseen domains. We study multi-source unsupervised domainadaptation, where labeled data are available from multiple source domains andonly unlabeled data are observed from the target domain. To address potentialdistribution shifts, we propose a novel Conditional Group DistributionallyRobust Optimization (CG-DRO) framework that learns a classifier by minimizingthe worst-case cross-entropy loss over the convex combinations of theconditional outcome distributions from sources domains. We develop an efficientMirror Prox algorithm for solving the minimax problem and employ a doublemachine learning procedure to estimate the risk function, ensuring that errorsin nuisance estimation contribute only at higher-order rates. We establish faststatistical convergence rates for the empirical CG-DRO estimator byconstructing two surrogate minimax optimization problems that serve astheoretical bridges. A distinguishing challenge for CG-DRO is the emergence ofnonstandard asymptotics: the empirical CG-DRO estimator may fail to converge toa standard limiting distribution due to boundary effects and systeminstability. To address this, we introduce a perturbation-based inferenceprocedure that enables uniformly valid inference, including confidence intervalconstruction and hypothesis testing.