Multi-step Bridged Refinement for classifying cross-domian data

Traditional approaches for classification require that the labeled data should have an identical distribution with the unlabeled data in order to build a reliable classifier. However, the identical distributed labeled data are often in short supply. In this situation, we may hope to borrow the labeled data in the relative domain to help the target task. To address this problem, we propose a transfer learning approach called Multi-step Bridged Refinement, which extends the Bridged Refinement algorithm. In the proposed method, we construct a series mixture models to bridge the source and target domain, through which the initial labels of the unlabeled data are refined towards the target distribution in a multi-step way. We empirically show that the proposed method can better discover the relatedness between domains and make better transfer. The result also shows the final accuracy is insensitive to the initial predicted labels as long as the refinement step is sufficient enough.