IGCRRN: Improved Graph Convolution Res-Recurrent Network for spatio-temporal dependence capturing and traffic flow prediction

Accurate traffic flow prediction is critical for traffic management and route guidance, enabling urban traffic to be free-flowing conditions and maximizing transport efficiency. In current prediction methods, the simple and fixed spatial graph only uses the prior knowledge of the traffic network, resulting in weak prediction performance. This paper proposes an Improved Graph Convolution Res-Recurrent Network (IGCRRN), which relies on uncertain spatio-temporal information for traffic flow prediction. In particular, a spatial dependence matrix that combines the origin graph matrix and the data-generated embedding node matrix is created. In this way, the spatial connection relationship can be obtained from the static graph information and changing traffic flow series, making the improved graph convolution block infer and quantify the different contributions in both spatial dependence and temporal dependence in a data-driven manner. In addition, the residual structure is employed to model the multi-level spatial dependence, and the IGCRRN-cell units based on the residual connection block and LSTM are designed to make the model automatically capture the spatio-temporal dependence in the traffic flow sequence. Experiments are conducted on two real traffic datasets, and the experiment results show that our proposed spatial dependence matrix can investigate the valuable information and consider the heterogeneity in the traffic flow. The IGCRRN model outperforms the baseline and state-of-the-art methods in prediction performance.

Automated summary

This paper proposes an Improved Graph Convolution Res-Recurrent Network (IGCRRN) that utilizes uncertain spatio-temporal information for accurate traffic flow prediction. By combining the origin graph matrix and the data-generated embedding node matrix, a spatial dependence matrix is created, allowing the graph convolution block to infer and quantify spatial and temporal dependence in a data-driven manner. Experimental results on real traffic datasets demonstrate that the IGCRRN model outperforms baseline and state-of-the-art methods in prediction performance.