Application of hybrid support vector Machine models in analysis of work zone crash injury severity

Abstract

Crash severity models are often used to analyze the adverse effects of highway work zones on traffic safety. In this study we evaluated application of hybrid support vector machine (SVM) and hyperparameter optimization models for improved accuracy of crash severity prediction. Two hybrid models were evaluated: a genetic algorithm-optimized SVM (GA-SVM) and greedy-search optimized SVM (GS-SVM) models. The dataset used in model development and testing contained 12,198 work-zone crash observations in New Jersey over three years, from 2016 to 2018. The results indicate that the GA-SVM model outperformed both GS-SVM and the SVM with default parameters in predicting the severity of work zone crashes. While GA-SVM provided the best accuracy, it had the highest computation time. Among more than dozen factors considered in the models, the findings suggest that crash type and posted speed limit were the most significant for estimation or prediction of work-zone crash severity. The modeling approach and methods demonstrated in this study can improve the accuracy of crash prediction models. Also, a two-stage sensitivity analysis was conducted to see the impact of associated factors based on the probability of crash severity in work zones. The key findings revealed that early morning, nighttime, rainy environmental condition, rear-end crashes, a roadway with no median, and a higher posted speed limit increased the likelihood of injury and fatality in the work zone areas. This improvement will in turn lead to better informed decisions about planning and implementing work zone safety enhancements aimed at reducing severity of crashes.