New Graph Neural Network Improves Traffic Flow Prediction

This paper introduces the Structure-Guided Spatiotemporal Attention Graph Neural Network (SGSAN), a novel approach designed to enhance traffic flow prediction. While existing deep spatiotemporal models excel in predictive accuracy, their lack of transparency hinders deployment in critical urban systems. SGSAN addresses this by explicitly learning a static Directed Dependency Graph (DDG) to identify stable, macroscopic traffic propagation paths. The model integrates an InfoNCE-based soft-coupling mechanism that anchors its dynamic spatiotemporal attention to this structural prior. This alignment provides a mechanistic explanation for the model's decisions, improving interpretability and trustworthiness, while also ensuring robust forecasting by preventing over-reliance on transient local noise. Furthermore, a two-stage optimization framework is employed to balance structural discovery with predictive accuracy. Experiments on real-world datasets confirm SGSAN's state-of-the-art predictive performance and its built-in interpretability, which aligns with the physical logic of traffic networks.