Prediction on traffic accidents severity levels leveraging machine learning-based methods on imbalanced data

Abstract

Traffic accidents are a significant problem in many countries, resulting in thousands of injuries and deaths every year. By estimating the severity of traffic accidents, traffic safety together with the crash survival rates could be improved, by taking effective prevention measures at the location where accidents are plentiful and severe. This paper studies the prediction by different classification methods on traffic accident severity levels. The data set used includes 1.6 million traffic accidents recorded in the United Kingdom, ranging from 2000 to 2016. It is a difficult task, since the levels are imbalanced distributed, making it difficult to classify the records accordingly. To tackle this problem, this work compared several classification methods on the task and evaluates their performances from the aspects of time, accuracy, and adaptability on imbalanced data sets. Experiments suggest that among the methods, the decision tree is the most recommended. This paper also provides suggestions for improvements on similar tasks.