Evaluating Traffic Safety and Geometric Characteristics Using Machine Learning Ensemble Techniques: A Case Study of Egyptian Rural Multi-Lane Divided Roads

Crash prediction models are essential for evaluating traffic safety by analyzing crash occurrence, frequency, or severity. Recently, machine learning techniques have gained prominence in statistical regression modeling and data analysis. This study assesses the effectiveness of machine learning in predicting crashes on Egyptian rural multi-lane divided roads using limited regional data. Supervised machine learning ensemble techniques were applied to predict crash-prone segments (classification) and estimate the total number of crashes per segment (regression). A comparative analysis aims to identify the most suitable method. The Synthetic Minority Oversampling Technique (SMOTE) addressed data imbalance, while K-means Clustering (KC) enhanced regression model accuracy, and the silhouette score assessed clustering quality. Feature importance analysis identified significant input variables for crash prediction. Input variables included traffic volumes, segment length, and road geometric characteristics. Reported crash data from 2008 to 2011 were analyzed using 177 samples (75% training, 25% testing). Classification model performance was assessed using accuracy, precision, F1-score, and recall, while regression models were evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and R² score. The Gradient Boosting (GBoost) algorithm achieved the highest F1-score and accuracy for classification, with 92% accuracy for training and 84% for testing, along with 93% precision and 84% recall. AdaBoost performed best for regression, with R² scores of 0.94 for training and 0.87 for testing. Classification models outperformed regression models in accuracy. Average Annual Daily Traffic was the most significant predictor of crashes, followed by section-length and pavement-width, while the number of U-turns was less significant.