Multi-scale Subgraph Contrastive Learning for Link Prediction

Existing works about link prediction rely mainly on pooling operations which cause loss of edge information or similarity assumptions, so that they are limited in specific networks, and mainly supervised learning methods. We propose a Multi-scale Subgraph Contrastive Learning (MSCL) method. To adapt to networks of different sizes and make direct use of edge information, MSCL converts a sampled subgraph centered on the target link into a line graph as a node-scale to represent links, and mines deep representations by combining two scales information, subgraph-scale and line graph node-scale. After learning the information of the two subgraphs separately by encoders, we use contrastive learning to balance the information of two scales to alleviate the over-reliance of the model on labels and enhance the model’s robustness. MSCL outperforms a set of state-of-the-art graph representation learning solutions on link prediction task in a variety of graphs including biology networks and social networks.