Hypergraph Neural Networks Based on Enclosing Subgraph Extraction for Temporal Link Prediction

The Graph Neural Network (GNN) methods based on enclosing subgraph extraction have achieved excellent results in static graph link prediction tasks. However, most real-world networks are dynamic and evolve over time; the traditional models cannot capture temporal evolution. Moreover, the existing temporal link prediction methods mainly use simple GNNs to learn spatial embeddings, which not only ignore the high-order relationships and local clustering structures between node, but also perform poorly in static graph link prediction tasks. In order to address the above problems, we propose Hypergraph for temporal link prediction (HGTLP) which is based on enclosing subgraph extraction. The proposed HGTLP constructs hypergraph through enclosing subgraph extraction and dual hypergraph transformation to learn the high-order spatial and temporal embeddings of each link to be predicted, and then combines the global knowledge of the collapsed graph to improve the prediction accuracy. Experimental results demonstrate that HGTLP achieves significant performance improvement compared to the state-of-the-art models in both link prediction and new link prediction tasks.