Abstract
This study applies machine-learning regression techniques to predict the frequency of pedestrian and bicycle crashes at urban intersections using roadways, traffic, and socioeconomic features. Five supervised machine learning models were used and evaluated using Percent Correct as the primary performance metric including: Elastic Net Linear Regression (LR), k-Nearest Neighbors (KNN), Support Vector Machine Regression (SVM), Decision Tree Regression (DT), and Random Forest Regression (RF). Elastic Net achieved the highest accuracy, correctly predicting crash counts at 78.2% of intersections, followed by KNN (76.0%) and SVR (75.1%). Random Forest showed moderate performance (73.9%), whereas Decision Tree regression performed the weakest (66.5%). SHAP analysis indicated that roadway functional class is the strongest predictor of crash frequency, with additional contributions from neighborhood demographic and housing characteristics. The findings demonstrate that compact, regularized models can outperform more complex structures and highlight the importance of integrating both roadway design and socioeconomic context when assessing vulnerable-road-user safety at intersections.