K-Nearest Neighbor Based Local Distribution Alignment

When massive labeled data are unavailable, domain adaptation can transfer knowledge from a different source domain. Many recent domain adaptation methods merely focus on extracting domain-invariant features via minimizing the global distribution divergence between domains while ignoring local distribution alignment. In order to solve the problem of incomplete distribution alignment, we propose a K-nearest neighbors based local distribution alignment method, where Maximum Mean Discrepancy (MMD) is adopted as the transfer loss function to reduce the global distribution discrepancy, and then a K-nearest neighbors based transfer loss function is also devised to minimize the local distribution difference for the complete alignment of source and target domain. The proposed method contributes to avoid the dilemma of incomplete alignment in MMD by local distribution alignment and improve its recognition accuracy. Experiments on multiple transfer learning datasets show that the proposed method performs comparatively well.