Multi-classification Prediction and Interaction Effects of Determinants for Accident Severity on Two-lane Highways in Plateau Mountainous Region

Current studies suffer from the insufficient prediction accuracy for multi-classification prediction and unclear interaction mechanisms for accidents severity on two-lane highways in plateau mountainous region. To address these issues, this study proposes an XGBoost-based three-classification prediction framework, optimized by a genetic algorithm (GA). The framework is tested based on accident data from 2012 to 2017 on mountainous two-lane highways in Yunnan. It integrates 14 features, such as road geometry, traffic environment, and type of involved vehicle. The model performance is compared with random forest (RF), support vector machine (SVM), and the baseline XGBoost model. Additionally, partial dependence plots (PDP) are used to explore the influence mechanisms of different risk determinants on accident severity. The results show that: ①The proposed GA-XGBoost model has the best overall prediction performance, with accuracy, precision, and recall rates reaching 81.57%, 73.12%, and 82.68%, respectively. After optimization with the GA algorithm, the predictive accuracy for injury and fatal accidents improves by 14.58% and 50.00%, respectively, compared to those of the pre-optimization model. The number of correctly classified fatal accidents is three times than that of the RF and SVM models. All these show significant improvement of the ability to predict severe accidents. ②Factors reflecting vehicle characteristics and traffic environment have a more significant impact on accident occurrence. Among them, the type of causing-trouble vehicle, type of involved vehicle, accident type, and daily traffic volume are the top four risk factors. ③Regardless of the type of accident, when pedestrians or motorcycles are involved, the severity of the accident is significantly increased. Among them, pedestrian involvement increases the severity of the accident by 1.25 to 5 times higher than that of any other involvement type. Additionally, as traffic volume increases, the impact of side collisions on accident severity gradually increases.