Continuous-Graph Attentional Neural Networks for Temporal Link Prediction

Abstract

Link prediction on temporal networks is a hot issue in the research of network evolution. Existing works typically employ graph neural networks and a temporal feature extractor to build prediction model. However, such methods are facing two problems: 1) the over smoothing becomes a challenge when considering capturing deeper spatiotemporal dependence. 2) temporal feature extraction is still a challenge. In this research, we introduce a novel link prediction model named Continuous-graph attentional neural networks for temporal link prediction (LP-CGA). The model is based on an improved Auto-Encoder, which not only embeds structure information of temporal networks but also considers the evolution trend. Then, the deeper spatiotemporal information is mined through an attention-based ordinary differential equation (ODE). Two real dynamic network datasets, ITC and Infocom06, are used for experiments. The experimental results show that the proposed model is more accurate compared to other baseline methods.