Joint feature extraction and instance re-weighting approach for unsupervised domain adaptation

Domain adaptation methods aims to learn a classifier on labeled data (source domain) and recognize data from a new domain with a different distribution (target domain). A lot of researches are carried out in this field to learning appropriate features or reweighing the instances to alleviate the difference between domain distributions. In this paper, we propose a unified framework paying attention to both feature extraction and instance reweighting, simultaneously. Two similar projection matrices are learned for two domains. Also, by scarifying the learned model the importance of the instances are adjusted, acceptably. The empirical results validate that the proposed method significantly improves classification accuracies compared to the state-of-the-art methods.