Rethinking Traffic Speed Prediction with Traffic Flow-Aware Graph Attention Networks

Accurate prediction of traffic states, particularly speed, is crucial for efficient traffic management and planning. In this context, advanced deep learning methods, such as Graph Neural Networks (GNNs), have become popular in traffic speed prediction tasks due to their ability to capture the spatial characteristics of the road network. The Graph Attention Network further developed GNN by leveraging the attention mechanism with GNN, enabling the consideration of the temporal property of network speed. However, existing GNN models for traffic prediction have certain limitations. These include the reliance on Euclidean distance-based adjacency matrices, issues with over-smoothing, and a lack of understanding regarding the physical meaning of attention values. In this study, we first introduce a novel approach called the traffic flow-aware adjacency matrix, which considers the actual traffic flow instead of relying on Euclidean distance. Additionally, we incorporate the Katz index into the adjacency matrix, enabling effective training with fewer layers and mitigating the over-smoothing problem. Furthermore, we explore the physical interpretation of attention values in the context of the network topology and traffic volume. Our findings show a positive relationship between attention values and closeness centrality, indicating that links with higher attention values have greater connectivity to other roads. Also, the normalized attention values increase during peak commuting hours, aligning with higher traffic volume periods. The result of our study proposed enhancements to improve the accuracy and interpretability of graph attention-based traffic prediction models. Future research can explore the scalability of the proposed approaches and features to more extensive road networks, such as the entire city of Seoul, to further validate their effectiveness and robustness.