Aligning Contrastive Clusters for Cross-Network Node Classification

Linyao Yang, Hongyang Chen, Xiao Wang

Abstract: Unsupervised domain adaptation (UDA) has shown promise in cross-network node classification, which involves learning a node classifier from a labeled source network to classify nodes in an unlabeled target network. However, existing methods tend to focus on matching the marginal distributions of the source network and the target network. This may not always lead to optimal performance since different networks may have different class-conditional distributions. In this paper, we propose a novel cross-network node classification framework named AlignCL, which can effectively reduce the expected classification error on the target network by exploiting the discriminative clustering information of the target network and aligning the class-conditional distributions of both networks. Technically, AlignCL utilizes a mix-hop graph encoder to learn node representations incorporating both local and global semantics and employs a contrastive learning framework to force the representations of both networks to form discriminative class-conditional clusters. Then, clusters of both networks are matched based on their optimal transport distances. Lastly, a class-conditional domain adaptation module is designed to align the corresponding clusters across domains. Extensive experiments demonstrate that AlignCL achieves state-of-the-art performance on several cross-network node classification tasks.