Crash Prediction on Road Segments Using Machine Learning Methods

Abstract

This study adopted the Highway Safety Information System’s (HSIS) data for crashes occurred on road segments to develop supervised machine learning prediction models. Five machine learning models are developed: Linear Regression (LR), Generalize Additive Model (GAM), Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN). A comparison among the five model was performed using the root mean square error (RMSE) and the mean absolute error (MAE) as quality model indicators. The results indicated that the RF model was found to produce the best crash prediction results. The findings suggested that the increase in Annual Average Daily Traffic (AADT) exponentially increased the number of crashes on highway segments. In addition, roadway segments with the higher design speed induced the lower number of crashes, compared to the segments with the lower design speed. For segments of shorter than 5-mile long, the number of crashes rapidly increased as the segment length increased. However, there was no substantial increase in the number of crashes as the segment length increased for segments of longer than 5 miles. Also, the greater number of lanes on a roadway segment, the greater chance for increasing the number of crashes. Finally, the moderate grades showed the highest risk for occurrences of crashes, respectively followed by flat and rolling grades. These findings are useful for transportation professionals to consider when designing highways.