Label-guided heterogeneous domain adaptation

Authors：Zhou, Zhiheng^*; Wang, Yifan; Niu, Chang; Shang, Junyuan

Source：MULTIMEDIA TOOLS AND APPLICATIONS, 2022, 81(14): 20105-20126.

DOI：10.1007/s11042-022-12483-1

• Heterogeneous domain adaptation
• Data distribution
• Distribution adaptation

Heterogeneous domain adaptation(HDA) mainly considers how to solve the target domain task with the help of the relevant knowledge of the source domain when both data distribution and feature space of the target domain are different from the source domain. In this paper, we propose Label-guided Heterogeneous Domain Adaptation method, which focuses on how to enhances the application of a small amount of labeled target domain data. In our algorithm, we consider learning a mapping matrix to map the data of two domains into a shared subspace and make predictions accordingly. Firstly, we match the marginal and conditional distribution of the source and target domain data. Secondly, considering the guidance of labeled data in the target domain, we combine all labeled data and adapted it to the unlabeled part of the target domain. Finally, we introduce F-norm to reduce the parameter complexity of the mapping matrix. We conduct extensive experiments on text-to-text and image-to-image transfer tasks, and the experimental results demonstrated that our algorithm is significantly superior to several state-of-the-art algorithms.