Dual-aligned unsupervised domain adaptation with graph convolutional networks

In recent years, graph convolutional networks have achieved great success in unsupervised domain adaptation task. Although these works make effort to reduce the distribution difference between domains, they do not take into account the issue of distribution difference reduction in the class level. In this paper, we propose a Dual-aligned Unsupervised Domain Adaptation with Graph Convolutional Networks (DUDA-GCN) framework to align domain distributions and the distributions of two domains corresponding to each category jointly. The framework contains two parts, i.e., a cross-domain feature extractor and a dual aligner of distribution. The former adopts a two-channel sub-network, with each channel to fully explore the relation among within-domain samples, based on GCN with shared weights to learn common feature representations of two domains. The dual aligner contains an adversarial domain discriminator and a category aligner, where the domain discriminator is designed to reduce the distribution difference across domains. A pseudo-label generator is designed to generate pseudo-labels for unlabeled samples. With the generated pseudo-labels of unlabeled samples and the real labels of labeled samples, the category aligner aligns the sample distributions across domains of the same category. Extensive empirical evaluation on three real-world datasets shows that DUDA-GCN can perform better than state-of-the-art related domain adaptation methods.