Classification Performance Evaluation of Traffic Accident Data Using Machine Learning

Road traffic accidents, which are a global problem, cause huge losses both in economic and social areas, while traffic accidents lead to casualties, injuries and death. According to the World Health Organization report, more than 1.25 million deaths occur each year as a result of traffic accidents, on the other hand, non-fatal accidents affect more than 20 million people. Although Great Britain has the world's safest road records, research shows that 5 people are killed every day in road traffic accidents. In order to identify the most effective factors related to accidents, researchers have developed and effectively used large data sets containing various information about previous accidents. In this academic study, using the recorded traffic accidents data of Great Britain, statistical models will be used to identify and classify the parameters causing traffic accidents. a detailed procedure of injury severity prediction using the Support Vector Machine, k-Nearest Neighbour and Gaussian Naïve Bayes classification techniques will be discussed. Furthermore, feature selection methods including Chi-square, Random forest, Support vector machine recursive feature elimination and Light gradient boosting machine, will be debated to identify the most important attribute of the traffic accidents. According to the latest available data set in 2018, traffic accidents data, accuracy rate of 77.40% was calculated with the k-Nearest Neighbour method, 78.98% with SVM-RBF and 77.71% with Gaussian Naïve Bayes. As a result of the classification for the severity of casualty, SVM-RBF and GNB often performed the best, giving the same result, at a rate of 87.80%. Classification for vehicle type, the best accuracy value in both test data and training data was obtained with SVM-RBF method with 84.53% and 84.36, respectively. While the percentage of accuracy in the KNN and GNB classification methods for the test phase was 82.20% and 83.33%, respectively, it was calculated as 82.24% and 82.95%, respectively, as a result of the analysis made for the training phase. Although there are close answers with three classification methods, SVM RBF classification shows a better performance than other classification tools.