Structure-Enhanced Graph Neural ODE Network for Temporal Link Prediction

Jinlin Hou, Xuan Guo, Jiye Liu, Jie Li, Lin Pan, Wenjun Wang

Abstract: Temporal link prediction aims to predict future links by learning the structural information and temporal evolution of a network. However, existing methods are heavily dependent on the latest snapshots, which hinders their power to reveal the essential evolutionary patterns based on and leverage them for dynamical inference. As a result, they generally achieve better predictions for the closest future snapshots than remote ones. Moreover, most methods do not take into account the effects of higher-order and global structure. To address these issues, we propose Structure-Enhanced Graph Neural Ordinary Differential Equation (SEGODE), a framework effectively performing dynamic inference by neural ordinary differential equation incorporating attention mechanisms and empowering to capture higher-order and global structure. To validate the proposed model, we conduct multiple experiments on a total of seven real datasets. The experimental results show that the SEGODE not only achieves good performance in link prediction but also maintains excellent results even under data-scarce conditions.