DAGCRN: Graph convolutional recurrent network for traffic forecasting with dynamic adjacency matrix.

Accurate and real-time traffic forecasting is of great significance for urban traffic planning, traffic control, and traffic management. However, the time-varying dynamic spatial relations and the complicated spatial–temporal dependencies are still open problems to be considered in traffic forecasting. To address these issues, we propose a graph convolutional recurrent network for traffic forecasting with a dynamic adjacency matrix within an encoder–decoder framework, named DAGCRN. The DAGCRN consists of a spatial relation extraction module (SREM), an adjacency matrix update module (AMUM), a dynamic graph convolutional recurrent module (DGCRM), and a global temporal attention module (GTAM). Specifically, SREM and AMUM are proposed to capture nodes' mutual relations at each time step and to model the evolution of the dynamic adjacency matrix, respectively. DGCRM captures the spatial–temporal dependencies of traffic data based on dynamic graph convolution and gated recurrent unit. GTAM is designed to extract the long-range temporal dependencies between future time steps and historical time steps. Extensive experiments on two real-world traffic speed datasets demonstrate that the proposed DAGCRN outperforms a number of representative baselines consistently. [ABSTRACT FROM AUTHOR]