Comparing Resampling Algorithms and Classifiers for Modeling Traffic Risk Prediction.

Area of Science:

• Traffic safety engineering
• Machine learning applications
• Road infrastructure analysis

Background:

• Road infrastructure significantly impacts traffic safety, necessitating further research.
• Predicting traffic crash risk remains challenging due to inherent uncertainties.
• Existing studies explore deep learning for traffic crash prediction.

Purpose of the Study:

• To develop and validate a robust traffic crash risk prediction model for road sections.
• To identify key features influencing traffic crash risk.
• To compare the performance of different machine learning classifiers for this task.

Main Methods:

• Collected real-world expressway traffic crash data (2013-2020).
• Classified road sections into low, medium, and high risk levels using the time-spatial density ratio (Pts).
• Employed data balancing algorithms (SMOTEENN) and ensemble classifiers (Random Forest) for model construction and analysis.
• Utilized partial dependence plots (PDPs) for feature importance analysis.

Main Results:

• Data balancing algorithms enhanced classifier performance.
• Ensemble classifiers demonstrated superior performance metrics.
• The combination of SMOTEENN and Random Forest achieved the highest classification accuracy.
• Feature analysis identified key road geometry, pavement, structure, and weather condition impacts.

Conclusions:

• The proposed traffic crash risk prediction method, particularly using SMOTEENN and Random Forest, is effective.
• The model provides valuable insights into factors influencing road safety.
• Future applications include road maintenance and design safety assessments.