Potential Collision Severity Prediction Based on Data Distribution‐Preserving Resampling

The majority of existing research on collision severity focuses on post‐collision severity, which is not conducive to collision prevention. This paper proposes a novel method for predicting the severity of potential collisions, aiming to establish a prediction model to predict the potential consequences of collisions before they occur, providing a basis for quantifying driving risk. In developing this model, two key challenges are addressed: how to effectively characterise the severity of potential collisions and how to manage the class imbalance caused by the scarcity of severe collisions. To tackle the first challenge, we introduce a systematic approach to find the most representative features of potential collision severity. For the second challenge, we propose a distribution‐preserving resampling method to address the class imbalance. This approach includes two techniques: Remove Redundant Under Sampling (RRUS) and Core Seed‐based Synthetic Minority Oversampling Technique (CS‐SMOTE), which transform the imbalanced dataset into a balanced one while preserving the distribution characteristics of the original dataset. Finally, using the National Highway Traffic Safety Administration (NHTSA) dataset and the XGBoost algorithm, a potential collision severity prediction model is developed. The results demonstrate that the model achieves a prediction accuracy of over 97.7%, outperforming comparison models developed using other classification algorithms.