Robust Traffic Prediction From Spatial-Temporal Data Based on Conditional Distribution Learning.

Area of Science:

• Intelligent Transportation Systems
• Machine Learning for Spatio-Temporal Data

Background:

• Traffic prediction is crucial for traffic management but challenged by complex spatial-temporal correlations.
• Existing graph neural network (GNN) models struggle with data distributions outside their training regions.

Purpose of the Study:

• To enhance traffic speed prediction accuracy by addressing limitations of current GNN models.
• To introduce a novel stagewise learning mechanism for improved spatial-temporal traffic data modeling.

Main Methods:

• Redefined speed prediction as a two-stage process: conditional distribution learning and speed regression.
• Introduced a mean-residue loss function with mean and residue loss components.
• Integrated a GNN architecture with the stagewise learning mechanism and mean-residue loss.

Main Results:

• The proposed method demonstrated superior performance on three public traffic datasets.
• Effectively handled complex spatial-temporal correlations in traffic speed data.
• Outperformed existing state-of-the-art traffic prediction methods.

Conclusions:

• The stagewise learning mechanism with mean-residue loss significantly improves traffic speed prediction.
• The approach enhances GNN capabilities for dynamic spatial-temporal modeling.
• This method offers a robust solution for intelligent traffic management systems.