Time-Aware Random Walk Diffusion to Improve Dynamic Graph Learning

How can we augment a dynamic graph for improving the performance of neural networks? Graph augmentation has been widely utilized to boost learning GNN-based models. However, most existing approaches only enhance spatial structure within an input static by transforming graph, and do not consider dynamics caused time such as temporal locality, i.e., recent edges are more influential than earlier ones, which remains challenging augmentation. In this work, propose TiaRa (Time-aware Random Walk Diffusion), novel diffusion-based method augmenting represented discrete-time sequence snapshots. For purpose, first design time-aware random walk proximity so that surfer along dimension well edges, resulting in spatially temporally localized scores. We then derive our diffusion matrices based on walk, show they become enhanced adjacency both localities augmented. Throughout extensive experiments, demonstrate effectively augments given leads significant improvements GNN models various datasets tasks.