Class-rebalanced wasserstein distance for multi-source domain adaptation

In the study of machine learning, multi-source domain adaptation (MSDA) handles multiple datasets which are collected from different distributions by using domain-invariant knowledge extraction. However, the current studies mainly employ features and raw labels on the joint space to perform domain alignment, neglecting the intrinsic structure of label distribution that can harm the performance of adaptation. Therefore, to make better use of label information when aligning joint feature-label distribution, we propose a rebalancing scheme, class-rebalanced Wasserstein distance (CRWD), for unsupervised MSDA under class-wise imbalance and data correlation. Based on the optimal transport for domain adaptation (OTDA) framework, CRWD mitigates the impact of the biased label structure by rectifying the Wasserstein mapping from source to target space. Technically, the class proportions are utilized to encourage distributional transportation between minor classes and principal components, which reweigh the optimal transport plan and reinforce the ground metric of Mahalanobis distance to better metricise the differences among domains. In addition, the scheme measures both inter-domain and intra-source discrepancies to enhance adaptation. Extensive experiments are conducted on various benchmarks, and the results prove that CRWD has competitive advantages.