Synchronous Spatiotemporal Graph Transformer: A New Framework for Traffic Data Prediction.

Area of Science:

• Computer Science
• Artificial Intelligence
• Transportation Engineering

Background:

• Modeling spatiotemporal relationships (STR) in traffic data is crucial but challenging for current graph networks.
• Existing methods often process temporal and spatial dimensions separately or use multiple local graphs, limiting long-term feature extraction and capturing complex relationships.

Purpose of the Study:

• To propose the Synchronous Spatio-Temporal grAph Transformer (S2TAT) network for efficient and comprehensive traffic data modeling.
• To address the limitations of existing methods in capturing nonlocal STR and handling data heterogeneity.

Main Methods:

• The S2TAT network integrates an attention mechanism and graph convolution within its core block to synchronously model nonlocal STR.
• It employs timewise graph convolution and a multihead mechanism to manage data heterogeneity.
• A novel attention-based output module captures valuable historical information, surpassing conventional average aggregation.

Main Results:

• The S2TAT network demonstrated superior performance on PeMS datasets.
• It achieved top-one accuracy, indicating highly effective spatiotemporal relationship modeling.
• The proposed method also exhibited reduced computational costs compared to state-of-the-art approaches.

Conclusions:

• The S2TAT network offers an efficient and effective solution for modeling spatiotemporal traffic data.
• Its integrated approach overcomes limitations of existing methods, providing enhanced accuracy and reduced computational load.