Distribution matching and structure preservation for domain adaptation

Abstract Cross-domain classification refers to completing the corresponding task in a target domain which lacks label information, by exploring useful knowledge related source but with different data distribution. Domain adaptation can deal such cross-domain classification, reducing divergence of domains and transferring relevant from target. To mine discriminant information samples geometric structure domains, thus improve performance, this paper proposes novel method involving distribution matching preservation for (DMSP). First, it aligns subspaces on Grassmann manifold; learns non-distorted embedded feature representations two domains. Second, space, empirical risk minimization regularization intra-domain graph is used learn an adaptive classifier, further adapting Finally, we perform extensive experiments widely datasets validate superiority DMSP. The average accuracy DMSP these highest compared several state-of-the-art methods.