a comparative study on performance of 4 machine learning algorithms in crash severity modeling

Articles in Press

Document Type : Scientific - Research

Authors

1 Department of transportation, faculty of civil engineering, Iran university of science and technology, Tehran, Iran

2 department of transportation engineering, faculty of civil engineering, Iran university of science and technology, Tehran, Iran

Abstract

Road crashes cause sociological and economical costs to societies, hence in recent years traffic safety studies has been a major concern. studies have been conducted in regards to crash severity modeling, but the majority of theses studies has traditionally used statistical modeling techniques, and the use of machine learning has been limited. Statistical methods have presupposed some conditions regarding variables and their relations, and if these conditions are violated, inference is affected. A wide spectrum of machine learning algorithm could be employed in crash safety modeling; hence the current study aims to comprehensively compare the performance of 4 widely used machine learning algorithms, namely classification and regression tree, artificial neural network, support vector machine and k-nearest neighbors. Results indicate that CART outperformed other models both in terms of accuracy and the prediction of the less populated crash severity level. Furthermore, the variable importance effected was analyzed using SHAP method, where at fault driver’s age, lighting condition and speed limit were found to be most significant across different models.

Keywords

Main Subjects

Quarterly Journal of Transportation Engineering

Articles in Press, Accepted Manuscript
Available Online from 11 October 2023

Files

History

  • Receive Date: 04 September 2023
  • Revise Date: 28 September 2023
  • Accept Date: 30 September 2023

Share

How to cite

Statistics