Spatial-temporal graph neural networks for traffic forecasting across multiple study sites

Abstract

Traffic forecasting has garnered research interest as a crucial technical component of adaptive traffic management over the last four decades. Since 2014, deep learning models have been actively proposed in this field of research, and graph neural networks have recently taken the major portion. Most studies, however, focus only on the accuracy of the models, and the robustness on various transportation networks and geospatial complexities is often neglected. To this end, this dissertation focuses on ensuring both the accuracy and robustness of traffic forecasting models based on graph neural networks. Techniques to effectively incorporate static structural characteristics depending on the geospatial complexity, and to reflect dynamic time-varying spatial correlation through online traffic data are studied. Also, a comprehensive assessment has been conducted to evaluate robustness and generalization capability of traffic forecasting models.