Improved Balanced Random Forest (iBRF): A flexible hybrid resampling-bagging framework for crash severity...

Class imbalance presents a notable challenge in accurately classifying severe crash outcomes, which typically constitute the minority class. Ignoring class imbalance may lead to biased or misleading interpretations. Moreover, relying on traditional resampling techniques results in potential issues like overfitting, and information loss. Hence, this study proposes the Improved Balanced Random Forest (iBRF), a novel ensemble framework designed to combine various resampling techniques with ensemble learning to derive a model that addresses the challenge of class imbalance.

The iBRF framework utilizes all minority cases and bootstrapped samples of the majority. SMOTE augments additional minority‑class instances, while the Neighborhood Cleaning Rule (NCR) and Random Undersampling (RUS) remove majority‑class instances to reduce noise and class overlap. The balanced data is utilized to train a decision tree in each iteration. Finally, a weighted soft voting based on the G-mean of each tree generates the final predictions. The performance of iBRF is tested using G-mean, and Matthews Correlation Coefficient (MCC) on holdout data (30%) and compared against other machine learning models designed for imbalanced datasets.

iBRF achieves better performance than other resampling (SMOTE, RUS, NCR, and CTGANs) and state-of-the-art ensemble learning approaches (XGBoost, RF, LightGBM, RUSBoost, SMOTEBagging, SMOTEBoost, UnderBagging, OverBagging, and OverBoost) based on balanced metrics. The iBRF algorithm substantially improves the absolute G‑Mean of the simple RF and SMOTE‑RF models by 9.30 and 4.50 percentage points, respectively.

The iBRF algorithm not only enhances classification performance but also helps identify risk factors associated with the minority class, which almost always correspond to severe‑outcome crashes and the focus of studies.