Prediction of traffic accident severity based on ordered classification models

Road traffic accidents remain a critical global issue, necessitating accurate prediction of injury severity to enhance safety management. Unlike most existing studies that treat accident severity as a nominal categorical variable, this research emphasizes its inherent ordinal nature (slight, serious, and fatal injury). So we primarily propose and evaluate two ordered classification (OC) models—Generalized Ordered Logistic Regression (GLR) model and a novel Weighted K-Nearest Neighbors (WKNN) classifier, which explicitly incorporate the order of injury severity. The GLR model utilizes the cumulative probability of categories instead of the probability corresponding to a single category in Logistic Regression (LR). The novel WKNN classifier assigns different weights to varying severity levels, thereby breaking the equality relationship between categories in the K-Nearest Neighbors (KNN) classifier. Using a real-world dataset of 12,316 traffic accident records from Addis Ababa (2017–2020) and combining Adaptive Synthetic Sampling (ADASYN) resampling. Furthermore, it is found that the predictive capabilities of the model are similar across the two resampling methods (ADASYN and Synthetic Minority Oversampling Technique (SMOTE)) and the two missing value handling approaches (deletion and imputation). Additionally, beyond the four unordered classification (UC) evaluation metrics, this paper employs an indicator named Total misclassification Cost (TC) that simultaneously accounts for the ordinal nature of the dependent variable and class imbalance. The TC value of the OC model can decrease by approximately 10% to 80%, significantly reducing the misclassification cost of the OC model. This study provides a structured and data-aware framework for traffic accident severity prediction, offering more reliable decision support for traffic safety management.