Optuna-Optimized Machine Learning for Traffic Accident Severity Prediction in Jordan under Class Imbalance Conditions
Optuna-Optimized Machine Learning for Traffic Accident Severity Prediction in Jordan under Class Imbalance Conditions

Summary/Abstract: A machine learning framework for predicting the traffic accident severity under class imbalance conditions is predented in Amman, Jordan. The methodology began with a preprocessing pipeline consisting of IQR-based (Interquartile range) outlier removal, label encoding, feature scaling, and Synthetic Minority Over-sampling Technique (SMOTE). Sixteen model configurations were considered, comprising four ensemble learning algorithms (XGBoost, LightGBM, Random Forest, CatBoost) under oversampling, with Optuna-based Bayesian hyperparameter optimisation. The result indicated that the class balancing using SMOTE had a greater influence on the predictive performance than hyperparameter optimisation, specifically for CatBoost model. Feature importance showed that vehicles involved, collision type, and speed are the strongest predictors of accident severity.

• Details
• Contents