DTS-AdapSTNet: an adaptive spatiotemporal neural networks for traffic prediction with multi-graph fusion

Traffic prediction is of vital importance in intelligent transportation systems. It enables efficient route planning, congestion avoidance, and reduction of travel time, etc. However, accurate road traffic prediction is challenging due to the complex spatio-temporal dependencies within the traffic network. Establishing and learning spatial dependencies are pivotal for accurate traffic prediction. Unfortunately, many existing methods for capturing spatial dependencies consider only single relationships, disregarding potential temporal and spatial correlations within the traffic network. Moreover, the end-to-end training methods often lack control over the training direction during graph learning. Additionally, existing traffic forecasting methods often fail to integrate multiple traffic data sources effectively, which affects prediction accuracy adversely. In order to capture the spatiotemporal dependencies of the traffic network accurately, a novel traffic prediction framework, Adaptive Spatio-Temporal Graph Neural Network based on Multi-graph Fusion (DTS-AdapSTNet), is proposed. Firstly, in order to better extract the hidden spatial dependencies, a method for fusing multiple factors is designed, which includes the distance relationship, transfer relationship and same-road segment relationship of traffic data. Secondly, an adaptive learning method is proposed, which can control the learning direction of parameters better by the adaptive matrix generation module and traffic prediction module. Thirdly, an improved loss function is designed for training processes and a multi-matrix fusion module is designed to perform weighted fusion of the learned matrices, updating the spatial adjacency matrix continuously, which fuses as much traffic information as possible for more accurate traffic prediction. Finally, experimental results using two large real-world datasets demonstrate that the DTS-AdapSTNet model outperforms other baseline models in terms of mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) when forecasting traffic speed one hour ahead. On average, it achieves reductions of 12.4%, 9.8% and 16.1%, respectively. Moreover, the ablation study validates the effectiveness of the individual modules of DTS-AdapSTNet.