MuDiS-GDA: Multiscale discriminative graph domain adaptation.

Abstract

Recently, graph domain adaptation (GDA) has received significant attention. The goal is to transfer knowledge from a labelled source graph to an unlabelled target graph to alleviate the lack of high-quality labels in the target domain. However, current GDA methods typically adopt a unified adaptation strategy for different scales of graph structures, which overlooks the divergences of domain shifts at different scales. Their adaptation strategies may cause structural misalignment and result in negative knowledge transfer. To address this issue, this study proposes a multiscale GDA method that learns and aligns multiscale discriminative features, which not only designs separate adaptation strategies for different scales of structures but also refines the discriminative information from multiscale structures to aid in classification. First, we propose a multiscale contrastive learning framework that constructs node-subgraph and node-graph contrasts to enhance the discriminative ability of node features from both source and target domains. The extracted discriminative information can be used to improve adaptation and finally facilitate the classification for the target domain. Second, we design specific domain adaptation methods for different scales based on the discriminative features. At the node level, we leverage the traditional adversarial domain adaptation strategy to learn domain-invariant node features with a domain classifier. At the subgraph level, we formulate a set of shared subgraphs based on the labels of the source domain, where each subgraph is associated with a class prototype. We then build a novel bidirectional matching loss between target nodes and shared prototypes, allowing nodes from both domains to align. Extensive experimental results on real-world datasets, compared to strong baselines, confirm the effectiveness of our method.