Multi-Spatiotemporal Dynamic Graph Convolutional Recurrent Network for Traffic Prediction

In the traffic flow prediction problem, most of the work uses a static adjacency matrix defined by prior knowledge to model the spatiotemporal correlation, ignoring the dynamic nature of traffic data. To accurately describe the dynamic relationships among nodes, a multi-spatiotemporal dynamic graph convolutional recurrent network is proposed for traffic flow prediction. First, the node embeddings are generated into new dynamic graph embeddings using a spatiotemporal embedding generator, and the node similarity is used to construct a dynamic graph, and the dynamic graph convolution module is used to fuse the dynamic spatial correlations at different times fully. Then, the current traffic signals and the dynamic graph embeddings are modeled with multi-temporal and spatiotemporal dependencies using the dynamic graph method. Finally, the obtained hidden representations are inputted into the prediction layer to obtain the final prediction results. Experiments are conducted on two real datasets, and the results show that the proposed model exhibits superior prediction performance compared to other baseline models.