A Unified Weighted MMD For Unsupervised Domain Adaptation

Abstract

Unsupervised domain adaptation (UDA) recognizes unlabeled domain data by using the classifier learned from another domain. Previous works mainly focus on domain-level alignment that usually ignores the class-level information, resulting in the samples of different classes being too close to be classified correctly. To tackle this challenge, we design a unified weighted maximum mean discrepancy (MMD) metric method, that measures the differences in empirical distributions of two domains by calculating the weights of different sample pairs adaptively. The unified weighted MMD method is proposed which combines the class-level alignment with domain-level alignment, making full use of intra-domain, inter-domain, intra-class, and inter-class information with adaptive weights, and it is easy to implement. Experiment results demonstrate that our method can obtain superior results from two standard UDA datasets Office-31 and ImageCLEF-DA, compared with other UDA approaches.