A hybrid algorithm based on machine learning (LightGBM-Optuna) for road accident severity classification (case study: United States from 2016 to 2020)

With population growth accidents have become one of the most crucial problems in countries. In the analysis of traffic accidents, some parameters such as the rate and severity of the accident are extremely important. And there have always been disagreements between the police and experts in the analysis. In this paper, after receiving the desired data and performing the pre-processing phases by preparing a database containing 458 thousand records and 126 characteristics of the most effective input variables to predict the severity of accidents of 4 classes of U.S. roads between 2016 and 2020 by providing a combined machine learning approach called LightGBM-Optuna was developed in the python programming environment. Then, 67% of the data were used for training, and the other 33% were used to testing the algorithm. The results showed that after presenting the model by the LightGBM algorithm with the optimal value of hyperparameters determined by the Optuna library, estimated predictions with the time of implementation of training and testing stages less than 1 min, the value of evaluation metrics, including an accuracy of 0.68, roc-auc of 0.90, the precision of 0.67, recall of 0.68, f1-score of 0.67, can be used as a method with simple interpretation and the high ability for traffic personnel and other authorities have helped enable better accident management and traffic safety programs by traffic engineers for identifying areas with higher risks.