Neighborhood Extended Dynamic Graph Neural Network

Abstract

Representation learning on dynamic graphs has drawn much attention due to its ability to learn hidden relationships as well as capture temporal patterns in graphs. It can be applied to represent a broad spectrum of graph-based data like social networks and further use the learned representations to solve the downstream tasks including link prediction and edge classification. Although some approaches have been proposed for dynamic graphs in recent years, most of them paid little attention to the evolution of the entire graph topology, leading to the lack of global information of what happened in nodes' neighborhoods during their update. We propose NEDGNN, a novel Neighborhood Extended Dynamic Graph Neural Network on dynamic graphs represented as sequences of time-stamped events. We introduce a temporal attention propagation module to generate messages for n-hop neighbors through a self-attention mechanism, which can help disseminate information among nodes' n-hop neighbors. Besides, a FIFO message box module is also applied to gain some time efficiency. Due to the introduction of these modules, NEDGNN outperforms many state-of-the-art baselines in several tasks. We also perform a detailed ablation study to test the effectiveness and time cost of each module.