Label-Aware Hierarchical Contrastive Domain Adaptation for Cross-Network Node Classification

Abstract

Cross-network node classification (CNNC) has gained a great deal of attention recently, which aims to transfer the knowledge from a label-rich source network to accurately classify nodes for a different but related unlabeled target network. To tackle the problem of network shift, the existing CNNC algorithms combine graph neural networks (GNNs) and domain adaptation (DA) to solve the problem. However, GNNs are vulnerable to network structure noises, and the traditional DA methods mainly focus on matching the marginal distributions and cannot guarantee the alignment of the class-conditional distributions of different networks. To remedy these deficiencies, we propose a novel label-aware hierarchical contrastive domain adaptation (LHCDA) model to address CNNC. On one hand, we use multi-head graph attention network (GAT) to learn noise-resistant node embeddings. On the other hand, a label-aware hierarchical contrastive domain adaptation module is designed to align the class-conditional distributions across networks at both node-node level and node-class level. Since target labels are unavailable, we use K-means clustering to generate pseudo-labels and employ the prediction confidence to reduce the noises. Extensive experimental results on six CNNC tasks demonstrate that the proposed LHCDA model is superior than previous state-of-the-art CNNC methods.