Recently, there has been a growing interest in Cross-network node classification (CNNC) which involves classifying nodes in a target network with limited labels by leveraging knowledge from a source network with abundant labels. In order to tackle this problem, we propose a domain-adaptive message passing graph neural network (DM-GNN). This model combines graph neural network (GNN) with conditional adversarial domain adaptation to learn informative representations for node classification that can be transferred across networks.
To begin with, we construct a GNN encoder using dual feature extractors. This allows us to separate the learning of ego-embeddings and neighbor-embeddings, enabling us to capture both commonalities and differences between connected nodes.
Next, we introduce a label propagation node classifier, which refines the label prediction of each node by considering its own prediction as well as the predictions of its neighbors. This helps improve the accuracy of node classification.
Furthermore, we devise a label-aware propagation scheme specifically for the labeled source network. This scheme promotes propagation within the same class while avoiding propagation between different classes. As a result, we obtain source embeddings that are more discriminative in terms of labels.
Finally, we perform conditional adversarial domain adaptation to incorporate the refined class-label information from the neighborhood. This adaptation ensures better matching of the class-conditional distributions across different networks.
We compare our proposed DM-GNN with eleven state-of-the-art methods and demonstrate its effectiveness in node classification.