Review on application of graph neural network in traffic prediction

To seek effective ways to improve the performance of spatial-temporal computing tasks in traffic prediction, and explore the prospects and challenges of applying the graph neural network technology in traffic prediction, the development of traffic prediction methods was reviewed. The advantages and limitations of model-driven methods, statistical models, traditional machine learning methods, and deep learning methods were summarized. The compatibility between graph networks and traffic networks was explained. The methods for constructing graphs were summarized. The data used for traffic prediction were classified. The commonalities and differences between different traffic prediction tasks were analyzed. The graph neural network models commonly used for traffic prediction tasks were concluded, including the convolutional graph neural network, graph attention network, graph autoencoder, and graph spatial-temporal network. The main factors and spatial-temporal modules considered when the graph neural network model was applied to traffic prediction were analyzed. The performance of various traffic speed prediction methods was compared. The impacts of different components of the graph neural network framework on the prediction performance were analyzed. The challenges and opportunities faced by traffic prediction based on the graph neural network were discussed from multiple perspectives such as the data multi-sourcing, application diversity, multimodality, dynamicity, model interpretability, uncertainty, and small sample learning. Relevant suggestions for the development trend of graph neural network were proposed. Research results show that compared with the benchmark model that only considers the time correlation, the performance of the prediction method based on the graph neural network improves significantly. The multi-mode time correlation, spatial-temporal attention mechanism, edge features, and external data can all significantly affect the prediction performance. The graph neural network provides a powerful means for modeling the spatial-temporal correlation with complex dynamicity of traffic networks. Currently, diversified models have been developed for the traffic state prediction problems. Future research can focus on developing efficient and dynamic spatial-temporal module integration architectures, designing modules that effectively integrate external data, expanding diversified application tasks, realizing the multi-mode traffic synchronous prediction, and developing efficient, reliable, and easy-to-explain models to achieve a balanced improvement in prediction accuracy and efficiency, so as to develop higher-level intelligent traffic. © 2023 Chang'an University. All rights reserved.