Subgraph Sampling and Common Neighbors Fusion-based Graph Neural Networks for Link Prediction

Link prediction using graph neural network is a critical issue in the field of graph-structured data mining. While many models employ message passing based on rooted subtrees to aggregate the neighboring structure information surrounding the target nodes, leading to ignore the connection information among these nodes. Moreover, combining local structural features with message passing enhance the efficiency of link prediction but fails to effectively discerning common neighbors with varying features. To address the above challenges, a new model called Subgraph Sampling and Common Neighbors Fusion-based Graph Neural Networks for Link Prediction (SSCNF-GNN) is proposed. Specifically, we construct a subgraph by sampling neighboring nodes of the target node for message passing. To discern common neighbors’ distinct features while incorporating local structural features into MPNN for link prediction, their node representations are combined with those of the target nodes as a link predictor. Extensive experiments conducted on real datasets show that SSCNF-GNN outperforms baseline methods. Ablation experiments confirm the essential roles of message passing on subgraphs and capturing diverse common neighbor features as local structural features.