Abstract

We propose a new general Graph Adversarial Domain Adaptation (GADA) based on semantic knowledge  reasoning of class structure for solving the problem of transferring from the source domain with non-shared and imbalanced classes to  target imbalanced domain, where non-shared classes mean the label space out of the target domain. Our goal is to leverage priori hierarchy knowledge to
enhance domain adversarial aligned feature representation with graph reasoning. In this paper, to address the sparse target classes and sparse  shared source classes in unsupervised domain adaptation, we equip  adversarial domain adaptation with Hierarchy Graph Reasoning (HGR) layer. For adversarial domain  adaptation, our HGR layer can  aggregate local feature to hierarchy graph nodes by node prediction and enhance domain adversarial aligned feature with hierarchy graph reasoning of all source classes. Our HGR contributes to learn direct semantic patterns by hierarchy attention, non-linear mapping and graph  normalization. Experiments on two benchmark datasets show our GADA methods consistently improve the state-of-the-art adversarial domain adaptation algorithms. The code is available at   https://gadatransfer.wixsite.com/gada.