Machine Learning Models for Proactive Road Safety: Evaluating Regression and Ensemble Techniques Based on Road Geometry

The complex interplay of factors contributing to road accidents necessitates the use of advanced predictive techniques capable of identifying the high-risk zones before incidents occur. The present research addresses this need by using a dataset with 16 road characteristics, such as curve radius, entry/exit speeds, and sight distance, and developing several models for this task, namely traditional linear models, including Simple Linear Regression (SLR), Ridge, Lasso, ElasticNet, ensemble techniques, including Random Forest Regressor (RFR), gradient boosting, Support Vector Regressor (SVR), and Extreme Gradient Boosting (XGBoost), and advanced gradient boosting frameworks, like LightGBM, and CatBoost. Among these, SLR achieved the best performance, with a Root Mean Square Error (RMSE): 6.92, and R2: 0.94 on the test set, while XGBoost ranked highest among the ensemble methods (RMSE: 14.55, R2: 0.75). The feature importance analysis revealed that the superelevation e (%), entry speed V(entry) (km/h), and mid-section speed V(mid) (km/h) were the most significant predictors across the models. This analysis offers valuable insights for the transportation authorities to predict accident-prone areas and implement targeted safety measures to reduce the road accidents.