A Survey on Spatio-Temporal Graph Neural Networks for Traffic Forecasting

Traffic forecasting plays an essential role in modern intelligent transportation systems (ITS) since it is the fundamental component of a wide range of real-life applications like alleviating traffic congestion or public transportation planning. With the pervasive use of sensors deployed in the cities and the evolution of computing resources, it is necessary to carefully explore the information in traffic data to further facilitate traffic applications. However, the complex spatial-temporal dependency of traffic can hardly be fully captured by the classic statistical methods and traditional machine learning methods. Recently, many frameworks based on Graph Neural Networks (GNNs) have been proposed to make full use of the abundant spatial information preserved in graph-structured data, and plenty of emerging GNN frameworks have reached state-of-the-art performance. One key advantage of GNNs is their ability to model irregular graphs with local aggregations. Motivated by the success of GNNs achieved in traffic prediction tasks, we present a survey in this paper that provides a summary on GNN-based deep learning approaches in traffic prediction from multiple perspectives. In general, traffic prediction methods based on spatial information often depend on graph construction, spatial information extraction, and spatial-temporal information fusion. We classify current methods by their roles in these steps during the traffic prediction pipeline, giving the advantages of these methods when introducing the detailed methods. At the end, research gap and trend of this discipline of study is summarised to be provided as a guideline for interested researchers.