Enhancing real-time traffic risk prediction with a cost-sensitive learning approach

Area of Science:

• Transportation Engineering
• Data Science
• Machine Learning

Background:

• Real-time traffic risk prediction is crucial for proactive traffic safety management.
• Existing models often neglect misprediction costs and varying consequences, impacting reliability.
• Vehicle trajectory data offers rich information for traffic state and risk extraction.

Purpose of the Study:

• To develop a cost-sensitive learning framework for traffic risk prediction that accounts for misprediction costs.
• To refine traffic risk classification into four distinct levels.
• To improve the reliability and accuracy of real-time traffic risk predictions.

Main Methods:

• Utilized the NGSIM dataset for empirical data extraction.
• Extracted traffic state variables and risk data at 5-second intervals.
• Developed a cost-sensitive learning framework with misprediction costs integrated.
• Optimized cost coefficients using a Genetic Algorithm (GA).
• Integrated the framework with four baseline models to create four enhanced models.

Main Results:

• The proposed enhanced models consistently outperformed baseline models across various metrics.
• Significant improvements were observed in identifying high-risk traffic events.
• Computational efficiency remained suitable for real-time traffic management applications.
• Reliability analysis validated the robustness of the GA-based optimization.

Conclusions:

• The cost-sensitive learning framework effectively enhances real-time traffic risk prediction accuracy and reliability.
• The proposed models offer a more dependable approach to proactive traffic safety management.
• Genetic Algorithm optimization provides a robust method for calibrating cost coefficients.