Handling Imbalanced Data for Real-Time Crash Prediction: Application of Boosting and Sampling Techniques

Document Type



traffic models; prove instruments; intelligent transportation systems; smart buildings; regression analysis; traffic accidents; data collections; logistics

Identifier Data



ASCE Library

Publication Source

Journal of Transportation Engineering, Part A: Systems


With a growing number of intelligent transportation system sensors and the networkwide deployment of those across the nation’s roadway facilities, current research and practices should concentrate on more proactive safety strategies. In recent years, real-time traffic data collected from ITS sensors have been utilized to develop crash prediction models. Real-time crash prediction models can be used to identify hazardous traffic conditions that might cause a crash. This study aims to examine how employing data mining techniques that account for imbalanced data could improve the predictive capability of real-time crash prediction models. The term imbalanced data refers to a condition where the number of observations in each class is not equally distributed among the data set (noncrash cases outnumber crash cases). To decrease the within-class variation of imbalanced data, the data were split into two traffic-state data sets: free-flow speed (FFS) and congestion. Three models, including logistic regression as the baseline, random forest (RF) with random undersampling, and Adaptive Boosting (AdaBoost), were estimated with each data set. The results were compared with the models that were estimated using the complete set of data. Model comparisons indicated that all three models achieved significantly better predictive results with the congested and FFS data sets as opposed to the data set containing all crashes and that, while in some cases the results of the undersampled RF model were slightly better than those of AdaBoost, both models outperformed the logistic regression model. The results of this study demonstrated that using models to deal with imbalanced data and lowering the variation of imbalanced data could substantially improve crash prediction accuracy. The findings could help traffic agencies to practically implement and deploy crash prediction models for real-time applications and develop crash prevention strategies accordingly.