RT-GCN: Gaussian-based spatiotemporal graph convolutional network for robust traffic prediction

• Published in: Information fusion 2024-02, Vol.102, p.102078, Article 102078
• Main Authors: Liu, Yutian, Rasouli, Soora, Wong, Melvin, Feng, Tao, Huang, Tianjin
• Format: Article
• Language: English
• Subjects: Gaussian distribution; Graph convolutional network; Missing data; Robustness; Traffic prediction
• ISSN: 1566-2535; 1872-6305
• DOI: 10.1016/j.inffus.2023.102078

Traffic forecasting plays a critical role in intelligent transportation systems (ITS) in smart cities. Travellers as well as urban managers rely on reliable traffic information to make their decisions for route choice and traffic management. However, noisy or missing traffic data poses a problem for accurate and robust traffic forecasting. While data-driven models such as deep neural networks can achieve high prediction accuracy with complete datasets, sensor malfunctions, and environmental effects degrade the performance of such models, as these models rely heavily on precise traffic measurements for model training and estimation. Consequently, incomplete traffic data poses a challenge for robust model design that can make accurate traffic forecasts with noisy/missing data. This research proposes the Robust Spatiotemporal Graph Convolutional Network (RT-GCN), a traffic prediction model that handles noise perturbations and missing data using a Gaussian distributed node representation and a variance based attention mechanism. Through experiments conducted on four real-world traffic datasets using diverse noisy and missing scenarios, the proposed RT-GCN model has demonstrated its ability to handle noise perturbations and missing values and provide high accuracy prediction. •Gaussian distribution representation enhance inherent robustness of neural network.•Variance-based atention nechanisn reduce the propagation of perturbation.•Batch Randon Noise help inprove robustness during training phase.•Models are tested on both noisy and nissing datasets.•Diverse perturbation scenarios are considered in experinents.