A framework for real-time traffic risk prediction incorporating cost-sensitive learning and dynamic thresholds.

Area of Science:

• Traffic safety engineering
• Machine learning applications
• Data-driven risk assessment

Background:

• Real vehicle trajectory data is used for real-time traffic risk prediction.
• Existing methods overlook misclassification costs and varying consequences, impacting reliability.
• Traffic risk prediction requires improved accuracy and reliability for proactive safety management.

Purpose of the Study:

• To refine traffic risk classification into four levels (no, low, medium, high risks).
• To incorporate misclassification costs using cost-sensitive learning (CSL).
• To enhance multi-class prediction performance with dynamic thresholds (DTs) and address class imbalance.

Main Methods:

• Utilized real vehicle trajectory data from the HighD dataset.
• Integrated CSL and DTs with four baseline machine/deep learning models.
• Employed a genetic algorithm (GA) to optimize cost coefficients and thresholds.

Main Results:

• CSL-DTs-based models demonstrated superior performance in multi-class traffic risk prediction compared to baseline models.
• Computation time for the proposed models is suitable for real-time applications.
• Robustness analysis confirmed model stability and reliability of GA optimization.

Conclusions:

• The proposed CSL-DTs approach significantly enhances the reliability of real-time traffic risk prediction.
• Findings support the advancement of proactive traffic safety management strategies.
• The study provides a robust framework for accurate and cost-aware traffic risk assessment.