Dynamic Self-Mutual Correlated Graph Convolutional Network for Traffic Prediction

Precise predicting of traffic patterns is imperative to improve the functionality and efficiency of intelligent transportation systems. Currently, the spatiotemporal deep learning methods are among the most successful and promising approaches. However, the task of traffic prediction encounters the following challenges that must be addressed: 1) How to dynamically describe the inhomogeneity of different periods. 2) How to capture global dependencies caused by hidden factors. In this paper, we propose the Dynamic Self-Mutual Correlated Graph Convolutional Network(DSMCnet) to address these challenges. DSMCnet employs a sequence-to-sequence architecture, consisting of an encoder to learn historical traffic patterns and a decoder to make predictions. This framework extracts the mutual correlation contained in global data, retaining the effect caused by both hidden factors and quantified factors. It uses the dynamic convolution operator based on node state distance to get the inhomogeneity. Then it extracts the self correlation with weighted parameters. The cooperation of dynamic convolution and global dependencies mechanisms effectively improves the expressive ability of traffic patterns. We evaluate the model on two real-world road network traffic datasets. Our evaluation suggests that the proposed model is approximately 7 % -12 % and 5%-11 % improved compared to baseline methods in terms of MAE and RMSE metrics respectively.