Traffic Accident Density Prediction Considering Injury Severity Based on Random Forest and GAM

Abstract

Traffic accident density prediction is an significant topic of road safety. This study uses some data in the census tracts in California to develop traffic accident density models using two different modeling approaches. In this paper, the dependent variable was the density of traffic crash which is a measure of the relative distribution of traffic crashes. The random forest regression was used to predict crash density. But many machine learning models like this are like a black-box, which can’t find the causal relationship between features and targets. For many practical problems, it is often more important to explain why a phenomenon happens than improving the model’s predictability. As a result, differed from pure predictive study, the generalized additive model (GAM) was introduced which was an interpretable statistical model to achieve explanatory analysis. The GAM explores which independent variables will have an impact on response variables and the extent of their impact. Thus, the traffic accident density prediction of this practical problem has more credibility analysis.