In-Depth with Spatial-Temporal Graph Neural Networks for Traffic Forecasting: An Overview with Attention

Abstract

Traffic forecasting is a well-known time-series problem in Intelligent Transportation Systems (ITS). Numerous machine learning techniques can be applied to enhance the accuracy of traffic forecasting, including traditional approaches and powerful deep learning methods. With the advancement of deep learning, Graph Neural Networks (GNNs) strongly improved the results of traffic prediction. Most state-of-the-art methods combine the GNN models with the "attention" mechanism to enhance efficiency. In this study, we focus on clarifying the importance and the trend of using the multi-head self-attention mechanism to improve traffic forecast accuracy in some state-of-the-art models. Successively, we compare representative models with and without multi-head self-attention and evaluate their performance on the new dataset named SEATTLE. Finally, we analyze the experimental results and discuss several research directions and opportunities in traffic forecasting.