Abstract: To explore the influence degree of influencing factors on different accident categories, three factors including road conditions, weather conditions and traffic flow state were considered, grid search method was used to determine the optimal parameters of hyperparameters, an improved road traffic accident risk assessment model based on random forest was established to investigate whether there is an accident, whether it is a vehicle-vehicle collision or a human-vehicle collision, whether it is a injury accident or a death accident. To quantify the contribution of influencing factors to the results of accident risk evaluation, a method of explaining the cause of traffic accident risk based on SHAP was proposed. The accident data of Beijing Jingkai Expressway and South Sixth Ring Road were used to calibrate and test the parameters of the proposed model, and the results were compared with traditional random forest, logistic regression and support vector machine (SVM). Results show that the constructed model has the best performance in the study of human-vehicle collision risk, and the recall rate is improved by 30% , 40% and 40% , respectively, compared with the traditional random forest, logistic regression model and support vector machine model, with high test accuracy. In total traffic accidents and injury accidents, the performance of the model on accident risk assessment is ranked secondly, and the improvement is about 20% , 10% and 10% compared with the baseline models. Compared with logistic regression model, there is a 30% increase in vehicle-vehicle collision accidents. There was no significant increase in death accidents. In terms of risk causes, the model considers that the current lane space headway, time occupancy and precipitation have relatively 30% , 30% and 10% effects on the overall accident risk. In the subdivision accidents, precipitation is the leading factor, followed by the current lane space headway and time occupancy.