Interaction representation-based subspace learning for domain adaptation

Generally, Transfer Learning or Domain Adaptation are used to solve domain inconsistency, but the conventional domain adaptation methods mostly only consider local information and ignore global information, and only consider one-way data and ignore the possibility of two-way data. Therefore, in this paper, we proposed an interactive representation-based framework for domain adaptation. In the novel framework, two low-rank based interactive representation models are built on both of source and target domains, which can be used to better align distribution discrepancy. Then, a distance constraint is designed to model the subspace relationship between source and target domain. Finally, the label-based regression is jointly used to earn extra discrimination for classification. Experiments on a number of public databases demonstrate that our method has competitive performance among comparison methods.