Traffic accident severity prediction with ensemble learning methods

In this study, decision tree-based models are proposed for classification of traffic accident severity. Traffic accident severity is classified into three categories. The data set used in the study belongs to the province of Kayseri, Turkey. The data consists of urban traffic accident reports (23074 accidents) between 2013 and 2021. There are 39 variables in the data set. As a result of data preprocessing, 15 variables that are meaningful and can be used for the model in the data set were determined. Since the input variables of the model mainly contain categorical data, they were coded with pseudo-coding and a total of 93 input variables were obtained. In the studies, ensemble learning methods such as Random Forest, AdaBoost and MLP methods were used. F1 scores of these methods were found to be 91.72%, 91.27% and 88.95%, respectively. Feature importance levels were calculated for 15 variables used in the model. Gini index and decision trees were used while calculating the importance of the features. Driver fault (0.64) was found to have the most effect on traffic accident severity. This study focuses especially on urban traffic accidents. Urban traffic is crowded in terms of both vehicles and pedestrians. As a result of this, according to the findings obtained in this study, traffic accidents occurred mostly at the intersections with crowded urban areas.