A novel ensemble learning method for crash prediction using road geometric alignments and traffic data

Abstract

As an important part of traffic safety analysis, crash prediction models using road geometric alignments and traffic data (CPM-GAs) have been regarded as the most classic way and can be used in stages of road safety evaluation and road operating and management. To improve the predictive performance of tradition CPM-GAs and avoid the overfitting problem of machine learning algorithms, a framework of CPM-GA based on ensemble learning theory and a new ensemble rule for connecting traditional models and machine learning models were proposed in this study. Results of the ensemble learning CPM-GA show that (1) classification and regression tree (CART) is recommended for important variable selection procedure before applying support vector machine (SVM), (2) machine learning models outperformed traditional models significantly in aspects of model fitting and prediction accuracy but are unstable in the sensitivity tests, (3) the new proposed ensemble method of traditional model and machine learning model can effectively improve the accuracy of traditional CPM-GAs by 10%–16% and reduce the variance of machine learning CPM-GAs by 12%–36% simultaneously. Finally, the ensemble method presented in this article may shed light on more research of crash prediction models based on ensemble learning theory.