Adversarial Spatial-Temporal Graph Network for Traffic Speed Prediction with Missing Values

Abstract

Traffic prediction plays a crucial role in constructing intelligent transportation systems. However, the acquired traffic data is often accompanied by missing values due to sensor faults in the data collection stage or communication failures in the data transmission stage. The lack of traffic data will bring great difficulties to the traffic prediction task. Existing traffic prediction models usually rely on the intactness of the data in terms of both spatial and temporal dimensions. In this paper, we propose an Adversarial Spatial-Temporal Graph Network model named ASTGnet, for traffic speed prediction. Our model acts like a multi-task learning framework to predict traffic speed and simultaneously impute missing values. This method effectively reduces the error accumulation between the imputation task and the prediction task. Moreover, we add the adversarial perturbation on the traffic data hidden state to enhance the robustness of the data representation embedding. We evaluate our model on real-world traffic datasets, and experimental results show that our framework has better prediction performance on datasets with missing values than baselines.