Title:Dynamic Link Prediction Using Graph Representation Learning with Enhanced Structure and Temporal Information

Abstract:The links in many real networks are evolving with time. The task of dynamic link prediction is to use past connection histories to infer links of the network at a future time. How to effectively learn the temporal and structural pattern of the network dynamics is the key. In this paper, we propose a graph representation learning model based on enhanced structure and temporal information (GRL\_EnSAT). For structural information, we exploit a combination of a graph attention network (GAT) and a self-attention network to capture structural neighborhood. For temporal dynamics, we use a masked self-attention network to capture the dynamics in the link evolution. In this way, GRL\_EnSAT not only learns low-dimensional embedding vectors but also preserves the nonlinear dynamic feature of the evolving network. GRL\_EnSAT is evaluated on four real datasets, in which GRL\_EnSAT outperforms most advanced baselines. Benefiting from the dynamic self-attention mechanism, GRL\_EnSAT yields better performance than approaches based on recursive graph evolution modeling.