Accident Prediction Accuracy Assessment for Highway-Rail Grade Crossings Using Random Forest Algorithm Compared with Decision Tree

Abstract

ABSTRACT Safety is a major concern of transportation planners and engineers in their design of highway rail grade crossings (HRGCs). Safety agencies rely on prediction models to allocate their crossing safety improvement resources. The prediction accuracy performance of those models is under-researched. This paper performs model forecasting accuracy comparison analysis for a proposed random forest method. Compared with the decision tree, the random forest method is capable of improving unbalanced data forecasting performance because of its bootstrap characteristic, which is a common resampling method to handle imbalanced data. Data imbalance is frequently encountered in safety analysis, where the use of inadequate performance metrics, such as accuracy, and specificity, will lead to overestimated generalization results. That is because the model/classifiers tend to predict the dominant class, non-crash class, in the area of safety analysis. The proposed random forest method is evaluated by various prediction performance measurements and compared with the decision tree. Results show that the random forest method dramatically improves the prediction accuracy without providing additional false negative predictions or false positive predictions which are known as false alarms.