<ul><li>Graph neural networks (GNNs) excel at node classification but require labeled data, leading to interest in unsupervised domain adaptation (UDA) for graphs.</li><li>Existing UDA techniques for graphs do not fully consider GNNs' structure and do not perform well when label distribution shift exists among domains.</li><li>A novel framework is proposed in this paper that utilizes link prediction to connect nodes between source and target graphs, enhancing message-passing and adaptation.</li><li>The framework includes an identity-preserving learning objective to maintain discriminative information in the target graph, with promising results shown on real-world datasets.</li></ul>

Bridging Source and Target Domains via Link Prediction for Unsupervised Domain Adaptation on Graphs

Discover more