Dynamic Dual-Channel Asynchronous Graph Neural Network for Traffic Prediction with Missing Values
Yanwei Yu, Longxin Guo, Dongliang Chen, Guiyuan Jiang, Bin Wang, Yongxin Tong, Junyu Dong · Fundamental Research · 2026
Traffic prediction has gained significant attention due to its crucial role in advancing intelligent transportation systems. Most traffic prediction models rely on idealized complete-data assumptions, yet real-world data often contains missing values from sensor or communication failures. Existing traffic prediction methods fail to model robust spatio-temporal features under missing data conditions, and current data imputation techniques cannot adequately handle incomplete traffic data with complex spatio-temporal dependencies. To this end, we propose a Dynamic Dual-Channel Asynchronous Graph Neural Network (D 2 AGNN) for traffic prediction with missing values. Our model aims to explore a joint multi-task training that leverages the synergy of imputing missing data and traffic predictions to enhance the accuracy of traffic predictions under incomplete data. First, D 2 AGNNdesigns a Dual-Channel Asynchronous Graph Neural Network (DAGNN) to generate enhanced node representations, effectively capturing hidden information transfers between nodes, even in the presence of missing data. Secondly, we present a dynamic graph learning that uses these representations to construct a dynamic traffic graph at each time interval, capturing the evolving states of traffic. Finally, a ST-Block processes inherent and diffusion traffic data separately, significantly enhancing prediction accuracy. Extensive experiments on two real-world traffic datasets demonstrate that our model consistently outperforms state-of-the-art models, especially in scenarios with substantial missing data. Compared with SOTA baseline, D 2 AGNNachieves average improvements in MAE, RMSE and MAPE of 9.47%, 11.35%, and 9.76%, respectively, on PEMS-BAY, and 9.40%, 9.46%, and 11.22% on METR-LA.