Link prediction is a crucial task in network analysis that aims to predict missing or potential links between nodes, with applications spanning social sciences, biology, and computer science. State-of-the-art methods have successfully converted this problem into a binary graph classification task by extracting h-hop subgraph structures.

However, this approach blocks information flow outside of h-hop subgraphs and requires additional memory. To address these limitations, we propose an end-to-end link prediction graph neural network incorporating a contrastive learning component.

Specifically, we utilize cross-scale contrastive learning to entrench subgraph information by maximizing mutual information between h-hop subgraph information and node representations around the target link.

Without explicitly extracting subgraph structures, the proposed method can update node representation with global information while obviating the requirements for additional memory.

Extensive experimental results across both plain and attribute graphs demonstrate that our proposed method achieves consistently competitive performance, outperforming other state-of-the-art methods in most cases with satisfying computation cost and fast convergence.

Q. Yang, W. Wei, R. Zhang, X. Feng, SCS: Subgraph contrastive supervised neural network for link prediction, Information Sciences 719 (2025) 122482.