Integrating generalized linear mixed models and XGBoost for safety performance function development on urban...

Area of Science:

• Transportation Engineering
• Traffic Safety
• Machine Learning Applications

Background:

• Urban arterial corridors face high crash risks due to traffic volume, complex designs, and varied road user interactions.
• Accurate prediction of crash frequency is crucial for effective traffic safety management and investment prioritization.

Purpose of the Study:

• To develop and evaluate Safety Performance Functions (SPFs) for urban arterial segments in Southeast Michigan.
• To compare traditional statistical models with ensemble machine learning techniques for crash prediction.

Main Methods:

• Employed Poisson, Negative Binomial, and Generalized Linear Mixed Models (GLMM).
• Utilized ensemble machine learning: Random Forest and Extreme Gradient Boosting (XGBoost).
• Assessed model performance using 5-fold cross-validation, focusing on predictive accuracy (R², RMSE).

Main Results:

• XGBoost demonstrated superior predictive accuracy (R² = 0.835, RMSE = 28.65).
• Key predictors identified across models include Annual Average Daily Traffic (AADT), segment length, speed limit, and pavement condition.
• GLMM provided an interpretable, length-adjusted SPF formulation.

Conclusions:

• Machine learning, particularly XGBoost, offers robust tools for predicting urban arterial crash frequency.
• Findings support data-driven decision-making for identifying high-risk road segments and optimizing safety investments.
• Models are adaptable for other regions with local calibration and data availability.