Moment matching of joint distributions for unsupervised domain adaptation

Unsupervised Domain Adaptation (UDA) is designed to transfer acquired knowledge from the source domain to an unlabeled target domain. In this paper, we present a comprehensive approach that seamlessly addresses both source-available and source-free UDA by matching the joint distributions across domains, independent of the availability of source data. Our methodology introduces three innovative criteria to quantitatively assess the divergences between the source and target data, as well as between the source model hypothesis and target data. The criteria decide whether the predicted labels of the target hypothesis are affected by the other knowledge of both domains in the form of a precise formula, thereby enabling targeted supervision in UDA. We evaluate the effectiveness through 37 image and text classification tasks across four different datasets, comparing their performance against the state-of-the-art models. Experiments demonstrate that the proposed approaches obtain superior accuracies for most of the tasks, especially for the source-free setting, which still exceeds HOMDA 0.6% on Office and DRDA 1.5% on Office-Home, even without direct access to source data.