Deep transformer-based heterogeneous spatiotemporal graph learning for geographical traffic forecasting.

Area of Science:

• Geospatial Artificial Intelligence (GeoAI)
• Urban Transportation Planning
• Traffic Management

Background:

• Deep learning models show progress in geographical traffic forecasting.
• Existing models struggle with long-term temporal dependencies and heterogeneous dynamic spatial dependencies.

Purpose of the Study:

• To propose a novel deep transformer-based heterogeneous spatiotemporal graph learning model.
• To address limitations in capturing long-term temporal and dynamic spatial dependencies in traffic data.

Main Methods:

• Incorporated a temporal transformer for long-term temporal pattern recognition.
• Introduced adaptive normalized graph structures for dynamic spatial dependency modeling.
• Developed a heterogeneous spatiotemporal graph learning framework.

Main Results:

• The model effectively captures long-term temporal patterns without simple data fusion.
• Adaptive graph structures enabled modeling of dynamic spatial dependencies and heterogeneous relationships.
• Achieved state-of-the-art results on four primary public datasets.

Conclusions:

• The proposed model significantly improves geographical traffic forecasting.
• It offers a robust solution for urban transportation planning and GeoAI applications.
• Demonstrates superior performance in handling complex spatiotemporal traffic dynamics.