Enhancing Traffic Prediction via Spatial Multi-Granularity Co-Evolving Mechanism

Accurate traffic prediction is crucial for effective urban planning and congestion management, aiming to forecast future traffic conditions by analyzing historical and real-time data. Traditional deep learning approaches typically rely on a single spatial granularity of traffic network data, which limits their capacity to fully capture the intricate details of multi-granularity spatial information, leading to decreased prediction accuracy in complex real-world scenarios. More critically, existing methods have shown deficiencies in effectively capturing the co-evolving dynamics across different granular levels, a key aspect for enhancing predictive accuracy. In this work, we introduce a novel Spatial Multi-granularity Co-evolving Framework (SMCF) to tackle these challenges. This innovative framework strategically integrates multiple spatial representations—micro, meso, and macro—enabling a comprehensive encapsulation of traffic dynamics at varying granularities. Unique to our approach is a co-evolutionary strategy where data from coarser granularities guide the predictive modeling at finer scales, fostering a dynamic and synergistic interaction across different levels of data granularity. Our extensive evaluations of two real-world traffic datasets confirm that the SMCF outperforms existing baseline models and demonstrates superior performance in traffic prediction tasks. This enhanced accuracy indicates our framework's ability to model complex spatial interactions and its potential applicability in real-world traffic management and planning scenarios.