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

Area of Science:

• Intelligent Transportation Systems
• Spatio-temporal modeling
• Machine Learning

Background:

• Accurate urban traffic prediction is crucial for intelligent transportation systems but is challenged by limitations in existing spatio-temporal models.
• Current models struggle to jointly capture local-to-global traffic evolution and distinct frequency dynamics (low vs. high).

Purpose of the Study:

• To propose a novel Coarse-to-Fine spatio-temporal learning framework with a Local-Global differential Network (C2F-LGNet).
• To address the limitations of discrete-time formulations and uniform model structures in capturing complex traffic dynamics.

Main Methods:

• Developed a unified framework with a coarse learner (multilayer perceptrons) for low-frequency patterns and a fine learner for high-frequency variations.
• Integrated a local-global ordinary differential equation (ODE) block combining multi-scale convolutions, self-attention, and neural ODEs for continuous spatio-temporal evolution.
• Employed a coarse-to-fine learning strategy to progressively model traffic dynamics.

Main Results:

• C2F-LGNet demonstrated superior accuracy, robustness, and generalizability across diverse datasets (traffic speed, metro ridership, bike-sharing, taxi flow).
• The framework effectively captures both dominant low-frequency patterns and complex high-frequency fluctuations in traffic data.
• Outperformed existing state-of-the-art models in urban traffic prediction tasks.

Conclusions:

• The proposed C2F-LGNet framework offers a more effective approach to urban traffic prediction by progressively modeling spatio-temporal dynamics.
• The integration of coarse-to-fine learning and local-global differential modeling enhances prediction accuracy and robustness.
• Validates the effectiveness of the proposed methods for intelligent transportation systems.