Contrastive Domain-adaptive Graph Selective Self-training Network for Cross-network Edge Classification

The performance of graph neural networks (GNNs) can degrade significantly when trained on the graphs with noisy edges connecting nodes from different classes. To mitigate the negative effect of noisy edges, previous studies have largely focused on predicting the label agreement between node pairs within a single network. So far, predicting noisy edges across different networks remains largely underexplored. To bridge this gap, our work studies a novel problem of cross-network homophilous and heterophilous edge classification (CNHHEC), aiming to predict the label agreement of edges in an unlabeled target network by transferring the knowledge from a labeled source network. A novel Contrastive Domain-adaptive Graph Self-training Network (CDGSN) is proposed. Firstly, CDGSN learns node and edge embeddings end-to-end by a GNN encoder with adaptive edge weights during neighborhood aggregation. Secondly, to facilitate knowledge transfer across networks, CDGSN employs an adversarial domain adaptation module to align edge embeddings across networks, and also designs a novel contrastive domain adaptation module to conduct class-aware cross-network alignment of node embeddings. As a result, the intra-class domain divergence can be mitigated while the inter-class domain discrepancy can be enlarged to yield network-invariant and label-discriminative node and edge embeddings. Moreover, CDGSN designs a selective positive and negative pseudo-labeling strategy to assign positive (negative) pseudo-labels to the target nodes with extremely high (low) prediction confidence of belonging to each specific class. Such pseudo-labeled target nodes would be employed to iteratively re-train the model in a self-training manner, so as to obtain more reliable target pseudo-labels progressively to guide class-aware domain alignment. Extensive experiments on benchmark datasets demonstrate the effectiveness of the proposed CDGSN on the CNHHEC problem. The performance of CDGSN is robust against various types of edge embeddings. When adopting three operators to construct edge embeddings, the proposed CDGSN can improve the state-of-the-art method for CNHHEC by an average of 0.5%, 2.8%, and 10.8% in terms of AUC, and 1.3%, 3.0%, and 6.3% in terms of AP, respectively.