Predicting Traffic Accidents Severity Using Multiple Analytical Techniques

Abstract

Road accidents are a major world economic and social problem, as shown by the report of loss of lives and properties in many countries worldwide. Reporting indicated the number of fatalities from road accidents per year of about 1.35 million and 50 million injuries was recorded or an average of 3000 deaths/day and 30,000 injuries/ day. Furthermore, its consequences have an impact on economic and social conditions in terms of health care costs of injuries and disabilities. The objectives of this paper are to implement four modeling techniques, Logistic Regression, Naïve Bayes, Support Vector Machine (SVM), and Artificial Neural Networks (ANN), to predict accident severity and compare the performance of these models in terms of their prediction accuracy. More than 117,000 accident records with over 32 variables were retrieved from London. The results showed that the nonlinear SVM model outperformed other techniques in terms of performance with an accuracy of 78.32%. On the other hand, the linear SVM was the worst overall model with an accuracy of 69.27%. In terms of training time, a considerable difference was found between two groups of models: Logistic Regression, Naïve Bayes on one hand, and SVM and ANN on the other group. The former required a shorter training time (less than 10 min for each model), while the latter had training times between 20 to 70 min per model. Overall, the nonlinear SVM seems to perform the best in terms of accuracy, while Naïve Bayes is the best for fast prediction. This result can be beneficial for researchers and practitioners to predict accident severity levels and suggest improvements to traffic safety.