Cross-Scale Contrastive Learning with Subgraph and Line-Graph Views for link prediction

Link prediction aims to infer potential or missing connections from observed network structures and represents a fundamental problem in complex network analysis. In recent years, graph neural networks have achieved significant progress in this field. However, most existing methods ignore edge-level structural information and rely solely on single-scale subgraph representations, limiting their ability to capture complementary information across structural scales. To address these limitations, we propose Cross-Scale Contrastive Learning with Subgraph and Line-Graph Views for link prediction. Specifically, enclosing subgraphs centered on target node pairs are extracted and aggregated to construct multi-scale subgraphs. Both the enclosing and aggregated subgraphs are further transformed into line graphs, enabling joint representation learning at the node and edge levels to capture complementary structural information. Finally, a cross-scale contrastive learning strategy is introduced to integrate multi-source representations and improve discriminative performance. Extensive experiments on nine publicly available datasets demonstrate that CSLG consistently outperforms existing baseline methods in terms of prediction accuracy and generalization performance. Our code is available at https://github.com/yabingyao/CSLG4LinkPrediction.