Joint Matrix Factorization and Structure Preserving for Domain Adaptation

Domain adaptation aims to learn robust classifiers for the target domain by transferring knowledge from the labeled source domain. However, most of the existing studies emphasize learning domain-invariant feature representations by employing distribution alignment on the feature space, which ignores the influence of data noise and structure knowledge. To address these issues, we propose a new domain adaptation approach, which can effectively reduce the impact of data noise and simultaneously thoroughly exploit the manifold data structures to transfer discriminative knowledge. Specifically, we jointly model matrix factorization, maximum entropy, distribution alignment in a unified framework, which can effectively alleviate the negative transfer of the outliers. Furthermore, we devise graph dual regularization to thoroughly explore the intrinsic manifold data structures, which can significantly reduce structure discrepancy across domains. Experimental results on various domain adaptation tasks demonstrate the superiority of the proposed method.