Multi-scale contrastive learning via aggregated subgraph for link prediction
Abstract: Link prediction seeks to uncover potential or future connections within a network using structural or attribute information. Recently, Graph Neural Network (GNN)-based methods have attracted considerable attention for their effectiveness in link prediction. However, most GNN-based approaches focus solely on single-scale input graphs, which limits their ability to comprehensively capture network structure information. In this paper, multi-scale subgraphs are introduced as input graphs to obtain complementary network structures from different perspectives. Simultaneously, to obtain embedding vectors with better representational capacity, contrastive loss from self-supervised learning is incorporated for link prediction. Specifically, Multi-scale Contrastive learning framework based on Aggregated Subgraph (MCAS) is proposed for predicting missing links. Firstly, we construct enclosing subgraph by extracting neighbors of target nodes. By applying aggregation operation to these subgraphs, different granularities of multi-scale subgraphs are obtained. Secondly, encoders are used to learn information from multiple scales of subgraphs separately. Next, contrastive learning is employed to achieve information balance among the multi-scale subgraphs. Finally, the minimization of the loss allows us to improve the model’s robustness. Empirical evidence indicates that our approach excels state-of-the-art methods on nine datasets, including biological and citation networks. All source code is publicly available at: https://github.com/yabingyao/MCAS4LinkPrediction.
Loading