A Crash Severity Prediction Method Based On Ensemble Learning

In recent years,the rapid development of the road transportation network system has greatly facilitated people’s travel,while it also lead to a series of problems.Many complicated factors and conditions have caused frequent traffic accidents,which not only poses a threat to life safety,but also causes losses to the economy.The ability to establish accurate road traffic accident prediction models is the key to further exploring the risk factors related to accidents and improving road traffic safety,which can promote the improvement of the transportation system.There are many different risk factors related to different types of road accidents.The crash data used in this study was collected from the Washington State Department of Transportation(WSDOT).This paper selects 11 factors which are associated with environmental,driver,road and vehicle characteristics in single-vehicle collisions.Based on the most serious injuries caused by road traffic accidents,the dependent variable was considered with two classes: death/injury accidents and property-only accidents.With thinking of advantages of relevant research,a crash severity prediction model established based on ensemble learning.Machine learning methods are effectiveness and accuracy which have been widely used in the transportation field.An ensemble-based system can obtain better predictive performance than any of the constituent learning algorithms.So,this paper designs a predictive model which combined a multi-objective optimization algorithm based on the basic framework of NSGA-II with random forest(NSGA-II+RF).The model includes the construction of the based learner,the secondary learner and the process of multi-objective optimization to obtain the optimal solution.The corresponding sensitivity analysis method is proposed according to the structure of the prediction model to extract the relationships of input risk factors and output severity.Compared the established model with decision tree,BP neural network,RBF-SVM,RF and NSGA-II+MLR by sensitivity,specificity,accuracy and G-mean from the same data.The results reveal the good performance of NSGA-II+RF and indicate that the specificity,accuracy,G-mean and sensitivity are 0.7199,0.7486,0.7940 and 0.8758 respectively.Applying the sensitivity analysis on the predictive models,we can identify the prioritized importance of crash-related factors.The results indicate that week is the most important risk factor than the others,which reflect the relationship between periodicity of traffic flow and the level of injury severity.Driver behavior or status is also important and people should constrain their own behaviors while driving.