STNN: A Spatial-Temporal Graph Neural Network for Traffic Prediction

• Published in: 2021 IEEE 27th International Conference on Parallel and Distributed Systems (ICPADS) pp. 146 - 152
• Main Authors: Yin, Xueyan, Li, Feifan, Wu, Genze, Wang, Pengfei, Shen, Yanming, Qi, Heng, Yin, Baocai
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01-12-2021
• Subjects: attention mechanism; Computational modeling; Correlation; Predictive models; Roads; Traffic control; traffic prediction; Training; transformer; Transformers
• ISSN: 2690-5965
• DOI: 10.1109/ICPADS53394.2021.00024

Accurate traffic prediction is of great importance in Intelligent Transportation System. This problem is very challenging due to the complex spatial and long-range temporal dependencies. Existing models generally suffer two limitations: (1) GCN-based methods usually use a fixed Laplacian matrix to model spatial dependencies, without considering their dynamics; (2) RNN and its variants are only capable of modeling a limited-range temporal dependencies, resulting in significant information loss. In this paper, we propose a novel spatial-temporal graph neural network (STNN), an end-to-end solution for traffic prediction that simultaneously captures dynamic spatial and long-range temporal dependencies. Specifically, STNN first uses a spatial attention network to model complex and dynamic spatial correlations, without any expensive matrix operations or relying on predefined road network topologies. Second, a temporal transformer network is utilized to model long-range temporal dependencies across multiple time steps, which considers not only the recent segment, but also the periodic dependencies of historical data. Making full use of historical data can alleviate the difficulty of obtaining real-time data and improve the prediction accuracy. Experiments are conducted on two real-world traffic datasets, and the results verify the effectiveness of the proposed model, especially in long-term traffic prediction.