Spatial-Temporal Dilated and Graph Convolutional Network for traffic prediction

Accurate traffic prediction is an indispensable work for urban traffic planning, traffic control and traffic management. But it is quite challenging to model the complex spatiotemporal correlation of the traffic data. To tackle this problem, we propose a Spatial-Temporal Dilated and Graph Convolutional Network(STDGCN) where first order approximation graph convolution is employed to capture the spatial correlations and dilated convolution is selected to extract the temporal one. Besides, the periodicity of traffic data is considered by multiple components. Extensive experiments on two real-world traffic datasets show our model obtains a competitive performance on traffic flow prediction task.