Modeling of traffic accidents with machine learning and different statistical methods: The case of Ankara province

Due to the developing technology in the world, traffic problems have appeared subsequent to the increase in the number of vehicles, and therefore solutions for the increase in accidents with fatal injuries and damage have become essential. In this respect, prediction models on traffic accidents have gradually gained importance and have been started to be used frequently in predictive models created with machine learning. In this study, 86.694 data belonging to Ankara province between the years of 2013-2020 obtained from the Decedents of accidents with fatal injuries of the General Directorate of Police and the Traffic Department were examined and analyzed. Apart from the machine learning model, many results were obtained in order to reveal the general characteristics of the accidents in the province of Ankara within the statistical scope with an explanatory nature in the data set. While creating accident prediction models, Logistic Regression, KNN (K Nearest Neighbor Classification), Support Vector Machine (SVM), Decision Tree (DT), Random Forest (RF), Extreme Gradient Boost (XGBoost), and Light Gradient Increasing Machine (Lightgbm) algorithms were used. The data set was evaluated as the Classification Problem in terms of fatal accidents. Due to this approach, there was unblance in the data set in terms of the number of deaths. In this respect, modeling was established on the original form of the data set during the modeling phase, and then SMOTE process was administered due to the unbalance and the models were reestablished. The results related to the available processes were evaluated according to the Accuracy and Recall metrics. The results obtained were noticed to be compatible with the studies in the literature.