In this paper, we propose a novel method, Spectral Augmentation for Graph Domain Adaptation (\method{}), to solve the domain adaptation problem in graph node classification using graph neural networks (GNNs). While previous studies on domain adaptation based on GNNs have mainly focused on aligning the feature spaces of source and target domains, we present a novel approach to align the feature spaces of each class in the spectral domain. This is based on the observation that nodes belonging to the same category in different domains have similar features in the spectral domain, while different classes are quite different. In addition, we develop a dual-graph convolutional neural network to exploit local and global consistency, and facilitate cross-domain knowledge transfer using a domain classifier with an adversarial learning submodule. Experimental results on various public datasets demonstrate the effectiveness of the proposed method.
