Prediction of roadway accident frequencies: Count regressions versus machine learning models

Prediction of accident frequency based on trac and roadway characteristics has been a very signicant tool in the eld of trac management. The accident frequencies on 185 roadway segments of the city of Mashhad, Iran, for the year 2007, were used to develop accident prediction models. Negative Binomial Regression, Zero In ated Negative Binomial Regression, Support Vector Machine and Back-Propagation Neural Network models were used to t the accident data. Both tting and predicting abilities of the models were evaluated through computing error values. Results show that the NBR model is the most eective model for predicting the number of accidents because of its low prediction and tting error values. Although the BPNN model has high tting capability, it does not have the prediction ability of the NBR model. Furthermore, the NBR is easily able to develop and interpret the role of eective variables, in comparison with machine learning models which have a black-box form. Marginal eect values for the NBR and ZINBR models, and sensitivity analysis of the SVM and BPNN models, reveal that Volume to Capacity ratio (V=C), Vehicle- Kilometers Travelled (VKT) and roadway width are the most signicant variables. An increase in V=C and roadway width will decrease the number of accidents, however, an increase in VKT and permission to park on the right lane of the roadway can increase the crash frequency.