DADiffNet: Delay-aware diffusion networks with adaptive subgraphs for large scale traffic forecasting

Area of Science:

• Transportation Science
• Data Science
• Network Analysis

Background:

• Accurate traffic forecasting is crucial for urban areas with increasing traffic demand.
• Existing methods struggle to capture upstream traffic perturbations and their downstream propagation delays.

Purpose of the Study:

• To develop a novel network model for enhanced spatiotemporal traffic forecasting.
• To explicitly address the challenge of modeling delayed diffusion of traffic perturbations.

Main Methods:

• Proposes the Delay-Aware Diffusion Network (DADiffNet) that reformulates spatiotemporal coupling.
• Models traffic-flow increments and uses adaptive subgraph structures for efficient topology encoding.
• Employs temporal differences to capture high-frequency perturbations and propagation delays.

Main Results:

• DADiffNet consistently outperformed fifteen baseline methods across eight real-world datasets.
• Achieved average improvements of 8.04% in MAE, 7.65% in RMSE, and 8.40% in MAPE.
• Demonstrated reduced memory consumption, enhancing accuracy, efficiency, and interpretability.

Conclusions:

• DADiffNet offers a superior approach to large-scale traffic forecasting by explicitly modeling delays and perturbations.
• The model provides a better balance between predictive accuracy, computational efficiency, and interpretability.
• This advancement is vital for managing traffic in rapidly expanding urban environments.