Time-Evolving Graph Convolutional Recurrent Network for Traffic Prediction

Accurate traffic prediction is crucial to the construction of intelligent transportation systems. This task remains challenging because complicated and dynamic spatiotemporal dependency in networks. While various graph-based networks have been proposed for prediction, most them rely on predefined graphs from different views or static adaptive matrices. Some implicit dynamics inter-node may be neglected, which limits performance prediction. To address this problem make more accurate predictions, we propose a model named Time-Evolving Graph Convolution Recurrent Network (TEGCRN), takes advantage time-evolving graph convolution capture adaptively at time slots. Specifically, first tensor-composing method generate adjacency graphs. Based these distance-based graph, module with mix-hop operation applied extract comprehensive information. Then resulting integrated into Neural structure form an general predicting model. Experiments two real-world datasets demonstrate superiority TEGCRN over multiple competitive baseline models, especially short-term also verifies effectiveness capturing dependency.