Testing The Performance of Random Forest and Support Vector Machine Algorithms for Predicting Cyclist Casualty Severity

Abstract

Traditional statistical regression models for predicting casualty severity have fundamental limitations. Machine learning algorithms for classifications have started to be applied in severity analysis in order to relax the assumptions and provide better accuracy in the models. However, the performances of highly advised classification algorithms for predicting cyclist casualty severity, which particularly occurred at roundabouts, have not been investigated comprehensively. Therefore, the study in this paper developed classification models for cyclist casualty severity prediction by applying the highest two advised algorithms in the literature namely Random Forest and Support Vector Machine. The dataset included 439 cyclist casualties which were recorded at give-way roundabouts in the North East of England. The predictive variables were sociodemographic information about cyclists, weather conditions, behavior-related contributory factors, speed limit, and roundabout geometrical parameters. 70% of the records were randomly selected for the training stage and 30% were used for the testing in both Random Forest and Support Vector Machine algorithms. After training the algorithm, the testing results showed that the Random Forest algorithm predicted the outcomes with 88.6% classification accuracy. On the other hand, Support Vector Machine algorithm predicted the testing values with 84.73% classification accuracy. The algorithms misestimated 18 and 20 of the casualties in Random Forest and Support Vector Machine, respectively. The outcomes suggested that both Random Forest and Support Vector Machine algorithms were applicable for cyclist casualty severity prediction models with high performance.