Modeling Pedestrian Red Signal Violations at Signalized Intersections Using Machine Learning

Pedestrian red-signal violations at signalized intersections have serious safety implications, particularly in rapidly urbanizing cities of developing countries. This study develops machine learning models to predict and explain pedestrian red-signal violations using data collected at five signalized intersections in Lahore, Pakistan (N = 1,460 observations). The observed mean violation rate of 58.8% highlights the severity of non-compliance. Five machine learning algorithms were evaluated: Logistic Regression, Decision Tree, Extra Trees, Random Forest, and CatBoost. Model evaluation employed two complementary validation strategies to assess within-site explanatory performance and cross-site generalizability. A nested stratified cross-validation scheme was used to assess model performance within observed intersections, while a Leave-One-Intersection-Out (LOIO) approach evaluated generalization to unseen intersections. Hyperparameters were optimized using the Optuna Tree-structured Parzen Estimator (TPE) sampler. The Extra Trees classifier achieved the best performance under stratified cross-validation. SHAP (SHapley Additive exPlanations) analysis indicated that intersection identifiers captured a substantial portion of the variation in violation outcomes, followed by G/C ratio and number of lanes. The LOIO evaluation revealed a substantial generalization gap, highlighting strong intersection-specific behavioral patterns. Despite limited generalizability, the interpretable insights provide actionable guidance for traffic safety practitioners.