Does Data Balancing Play an Important Role in Classifying Road Accident Severity in Thailand?

Unlike many studies that focus on binary classification, this study uniquely addresses road crash severity (RCS) classification in Thailand using three categories. A dataset of 31,817 cases collected between 2021 and 2022 was analyzed using nine features and four classifiers: Multinomial Logistic Regression (MLR), Least Absolute Shrinkage and Selection Operator (LASSO), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). The severity of the road crash was divided into three classes: no injuries or fatalities (46.47%), injuries (42.97%), and fatalities (10.56%). To mitigate class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was implemented. Subsequently, all classification performances were improved, especially in terms of the F1-G-Mean. Among four classifiers, LASSO and XGBoost showed the highest G-Mean of 0.6928 and sensitivity of 0.4877, respectively. RF presented the highest Fl-Score of 0.4359 and Fl-G-Mean of 0.4189. At α = 0.05, the Friedman test confirmed the significant performance differences between imbalanced and SMOTE scenarios. However, the overall improvements were moderate, suggesting that the data balancing alone is insufficient for optimal performance. These findings highlight the critical role of addressing class imbalance in improving classification reliability while emphasizing the need for additional techniques, such as advanced resampling, feature engineering, and hyperparameter tuning, to further enhance model performance. The study provides valuable insights for developing data-driven strategies to reduce road accident impacts in Thailand.