Which approach better samples extreme traffic conflicts? Conventional- vs. machine learning-based sampling methods.

Area of Science:

• Transportation Engineering
• Traffic Safety Analysis
• Machine Learning Applications

Background:

• Extreme value theory is used for crash risk estimation via sampling extreme traffic conflicts.
• Current methods lack standardized procedures for threshold selection and severity alignment evaluation.

Purpose of the Study:

• To address limitations in conventional extreme value theory sampling for crash risk.
• To evaluate machine learning models for improved sampling of extreme traffic conflicts.
• To compare machine learning-based sampling with conventional methods for severity alignment.

Main Methods:

• Investigated autoencoder neural network and isolation forest machine learning models.
• Utilized a database of vehicle-to-pedestrian conflicts at urban signalized intersections.
• Compared machine learning sampling with a conventional baseline technique.

Main Results:

• Machine learning methods produced extreme conflicts that better aligned with conceptual severity levels.
• Isolation forest demonstrated superior preservation of empirical tail distribution characteristics.
• Machine learning-based sampling offers improved contextual representation for extreme value distribution modeling.

Conclusions:

• Machine learning sampling methods, particularly isolation forest, enhance the accuracy of extreme traffic conflict identification.
• These advanced techniques offer a more robust approach to crash risk assessment compared to conventional methods.
• The findings support the integration of machine learning into traffic safety analysis for proactive risk management.