TraverseNet: Unifying Space and Time in Message Passing for Traffic Forecasting

Area of Science:

• Artificial Intelligence
• Computer Science
• Data Science

Background:

• Traditional traffic forecasting models often analyze spatial and temporal dependencies separately, compromising spatial-temporal integrity.
• Existing methods overlook the dynamic and delayed nature of neighbor influences on traffic conditions.
• Accurate traffic prediction requires a unified approach to spatial-temporal data with topological structures.

Purpose of the Study:

• To propose TraverseNet, a novel spatial-temporal graph neural network designed to unify spatial and temporal dependencies in traffic data.
• To capture inner spatial-temporal dependencies by treating space and time as an inseparable whole.
• To exploit evolving spatial-temporal dependencies for each node using message traverse mechanisms.

Main Methods:

• Developed TraverseNet, a spatial-temporal graph neural network architecture.
• Implemented message traverse mechanisms to model evolving spatial-temporal dependencies.
• Treated space and time as a unified, inseparable entity for data analysis.

Main Results:

• TraverseNet effectively unifies spatial and temporal dependencies in non-Euclidean spaces.
• The model accurately captures inner spatial-temporal dependencies for traffic data.
• Ablation and parameter studies validated the effectiveness of the proposed TraverseNet.

Conclusions:

• TraverseNet offers a significant advancement in spatial-temporal traffic forecasting.
• The unified approach to space-time dependencies improves the modeling of traffic dynamics.
• The novel message traverse mechanism enhances the capture of evolving neighbor influences.