C2F-LGNet: A coarse-to-fine framework with local-global differential modeling for traffic prediction.

Abstract

Urban traffic prediction is essential for intelligent transportation systems, as it supports congestion mitigation, traffic management, and efficient urban mobility planning. However, accurate forecasting remains challenging because existing spatio-temporal models often rely on discrete-time formulations that are insufficient for jointly characterizing local-to-global traffic evolution. Moreover, models with uniform and complex structures overlook the fact that traffic dynamics contain both easily learnable low-frequency regularities and harder high-frequency fluctuations. To address these issues, we propose a Coarse-to-Fine spatio-temporal learning framework with a Local-Global differential Network (C2F-LGNet), which progressively models coarse- and fine-grained traffic dynamics in a unified framework. Specifically, traffic information is first processed by a coarse learner built upon multilayer perceptrons, which progressively captures dominant and stable low-frequency traffic patterns while preserving more complex residual information for subsequent fine-grained refinement. A fine learner is then developed around a local-global ordinary differential equation (ODE) block to further characterize complex high-frequency variations and refine the traffic representations from both local and global perspectives. The local-global ODE block combines multi-scale convolutions for local spatial correlations, multi-head self-attention for global interactions, and a neural ODE for the continuous evolution between local and global states. Through this coarse-to-fine learning process, C2F-LGNet achieves progressive modeling of traffic dynamics, resulting in more accurate and robust predictions. Extensive experiments on traffic speed, metro ridership, bike-sharing, and taxi flow datasets demonstrate that C2F-LGNet consistently outperforms state-of-the-art models in accuracy, robustness, and generalizability, validating the effectiveness of the proposed coarse-to-fine learning and local-global differential modeling.