Study of Supervised Learning Algorithms for the Prediction of Road Accident Severity in Senegal

Abstract

According to a literature search, current studies on road accidents in Senegal are mostly conducted using conventional descriptive statistics, which, however, does not correctly identify cause and effect relationships and is unable to establish models that could predict accidents. An alternative way to reduce traffic accidents is to develop a model for predicting accident fatality. This model relates the severity of accidents to the main actors involved, namely the driver, the vehicle and the various characteristics of the road environment. This paper presents a study of several models for predicting the severity of traffic accidents in Senegal based on supervised learning algorithms such as Random Forest, k-nearest neighbor, SVM, logistic regression and naive Bayesian classifier, in order to estimate the accident severity from historical data. To solve the problem of unbalanced classes, we use several measures of models' performance namely percentage performance (Accuracy), area under the ROC curve, accuracy, Recall and F1 criterion. The best results were obtained by the Random Forest and SVM algorithms respectively, based on the percentage performance and the F1 criterion. Regarding the value of the area under the ROC curve, the best results were obtained by Random Forests and K-nearest neighbor algorithms, respectively. The Random Forests algorithms, which give the best results in model practice, can be used in studies on the prediction of road accident severity.