Generic and Dynamic Graph Representation Learning for Crowd Flow Modeling

Many deep spatio-temporal learning methods have been proposed for crowd flow modeling in recent years. However, most of them focus on designing a spatial and temporal convolution mechanism to aggregate information from nearby nodes historical observations pre-defined prediction task. Different the existing research, this paper aims provide generic dynamic representation method modeling. The main idea our is maintain continuous-time each node, update representations all continuously according streaming observed data. Along line, particular encoder-decoder architecture proposed, where encoder converts newly happened transactions into timestamped message, then related are updated generated message. role decoder guide process by reconstructing based node representations. Moreover, number virtual added discover macro-level patterns also share among spatially-interacted stations. Experiments conducted two real-world datasets four popular tasks result demonstrates that could achieve better performance than baseline methods.