Quantitative Study of Traffic Accident Prediction Models: A Case Study of Virginia Accidents

Abstract

Traffic accidents are a serious problem that threatens people's lives, health, and properties. Thus, decreasing traffic accidents is a crucial demand for public safety. This paper proposes two data mining models to predict accident risks based on the decision tree and the naive Bayes algorithms. The purpose of the classifiers is to predict the potential severity of a traffic accident based on a set of data attributes related to the weather factors, accident timing, and properties of the road. The models are developed using data on accidents in Virginia between 2016 and 2021. Several metrics are considered to measure the performance of each model such as accuracy, precision, recall, and F1-score. Furthermore, to statistically compare the performance of the prediction models, the study employs three quantitative analysis tools, approximate visual test, paired observations, and ANOVA. The experimental results revealed that the decision tree outperforms naive Bayes in terms of prediction accuracy.