Pseudo-Label Validation for Unsupervised Domain Adaptation

Model selection---choosing among candidate predictors, hyperparameters, or checkpoints---is a critical yet underaddressed bottleneck in unsupervised domain adaptation. Without target labels, standard validation fails, and existing unsupervised criteria rely on structural assumptions specific to feature alignment or image classification. We propose to construct a surrogate target validation loss by pseudo-labeling unlabeled target points with an imputation model, enabling model selection via standard holdout evaluation on the target domain. The approach is agnostic to the adaptation method, requires no density-ratio estimation, and applies to any generalized linear model loss, including logistic and square loss. Our central theoretical contribution is an oracle inequality showing that pseudo-label bias---not variance---governs selection quality, mirroring the classical insight that holdout validation tolerates label noise. Experiments on kernel methods regularization selection and deep backbone fine-tuning confirm that surrogate validation reliably tracks target risk without any target label.