Interpretable machine learning for evaluating speed management strategies using driver behavior data, case study of Kerman, Iran

Speed limit signage, monetary fines, and demerit point systems are methods that can be used to reduce speed, yet their behavioral effectiveness across driver profiles remains underexplored. This study examines how socio-demographics, driving experience, violation histories, and risky behaviors shape drivers’ compliance with these three instruments. Using 136 valid questionnaire responses from Kerman Province, we defined three binary target variables representing compliance sensitivity. A broad set of supervised machine learning models was trained via 5‑fold cross‑validation, with SMOTEN applied exclusively to training data to prevent leakage. Performance was assessed using accuracy, precision, recall, F1‑score, and ROC‑AUC, with SHAP values quantifying feature contributions. For speed sign compliance, Elastic Net Logistic Regression performed best (test accuracy = 0.71, ROC‑AUC = 0.72) with minimal overfitting. For fines and demerit points, Histogram‑based Gradient Boosting yielded the most competitive performance (test accuracies of 0.61 and 0.64; ROC‑AUCs of 0.54 and 0.56, respectively). SHAP analyses revealed that severe speeding, mobile phone use, improper maneuvers, crash history, and extended driving hours consistently predicted non‑compliance, whereas education, number of children, and driving experience promoted greater compliance. These findings suggest that risky behaviors and prior violations are primarily associated with lower deterrence sensitivity, with socio‑demographic factors playing secondary, moderating roles. Given the cross‑sectional, self‑reported nature of the data, the regional sample (Kerman Province), and the moderate sample size, these results should be interpreted as exploratory associations requiring validation in larger, more diverse, and ideally longitudinal populations, rather than as direct evidence of enforcement effectiveness.