Spiking Neural Networks-Based Neighborhood Co-Occurrence Link Prediction

Link prediction is an essential downstream task in graph structure analysis, holding significant value in various application scenarios such as social networks and network security. As a result, research on link prediction has garnered widespread attention. Current models for link prediction tasks typically employ recurrent neural networks to learn the representations of nodes and edges in the graph structure for subsequent graph learning tasks. However, this approach faces challenges related to gradient descent/explosion and computational efficiency. Therefore, in this paper, aiming to reduce model training time and computational costs, we propose a neighborhood co-occurrence model based on spiking neural networks (SNNs) for link prediction tasks. We utilize SNNs to sequentially model the linkage relationships between nodes in the graph, facilitating aggregation operations for nodes and neighbors, updating neighborhood representations to better capture the evolution of dynamic graphs. Additionally, we construct structural features based on common neighbors among nodes to predict unknown or potential links, enabling faster completion of link prediction tasks. To validate the effectiveness of the model, we conduct experiments on five largescale real-world datasets. The experimental results demonstrate that our model outperforms baselines in both transductive and inductive link prediction tasks while significantly improving the efficiency of model execution.