An adaptive prediction model for randomly distributed traffic data in urban road networks

Effective and efficient traffic prediction can provide a reliable data basis for traffic management in Intelligent Transportation Systems (ITS). While various machine learning methods have been proposed to enhance prediction accuracy in recent decades, there remain potential issues to be further addressed. Firstly, the inherent randomness of traffic dynamics usually leads to some outliers in historical observations, which may deviate the model parameter estimation when utilizing deep learning-based models to learn data distribution. Secondly, the spatial correlation among the road sections may dynamically change over time, posing challenges for modeling. In addition, due to the complexity of urban traffic networks, capturing such non-linear spatial dependencies based on the global road structure may consume huge computational resources. To address these issues, this paper proposes an adaptive temporal graph attention network (ATGAN), which is implemented in two steps: (1) An outlier time series filter (OTSF) technique is introduced to mitigate the adverse impact of outlier points and to adaptively learn the distribution of fluctuations of traffic data; (2) We design a group attention temporal graph convolutional network (GA-TGCN) to model the spatiotemporal features among neighboring road sections, which is achieved by adjusting the spatial correlation matrix dynamically in each training epoch with attention mechanism. We evaluate the prediction performance of ATGAN on two real-world datasets and the results show that our model can achieve higher prediction accuracy in less computational time compared with baseline methods.