MVSTT: A Multiview Spatial-Temporal Transformer Network for Traffic-Flow Forecasting.

Area of Science:

• Artificial Intelligence
• Computer Science
• Transportation Engineering

Background:

• Traffic-flow prediction is vital but challenging due to complex spatial-temporal dependencies and unpredictable events.
• Existing methods often use single- or dual-view learning, limiting their ability to capture intricate spatial-temporal features.

Purpose of the Study:

• To propose a novel multiview spatial-temporal transformer (MVSTT) network for enhanced traffic-flow prediction.
• To effectively learn complex spatial-temporal domain correlations and patterns from multiple data views.

Main Methods:

• Developed a temporal view with short-range and long-range gated convolution components for multi-granularity temporal knowledge aggregation.
• Designed a dual-graph spatial learning module to capture both fixed and dynamic spatial dependencies and edge evolution.
• Integrated a spatial-temporal transformer for multiview knowledge fusion to mine diverse spatial-temporal features.

Main Results:

• The proposed MVSTT network demonstrated superior performance in learning complex spatial-temporal features.
• Extensive experiments on four real-world traffic datasets confirmed the method's effectiveness.
• MVSTT consistently outperformed state-of-the-art baseline methods.

Conclusions:

• The MVSTT network offers a robust solution for accurate traffic-flow prediction by leveraging multiview learning.
• The proposed architecture effectively addresses limitations of existing approaches in capturing complex spatial-temporal dynamics.