A comparative analysis of risk factors in taxi-related crashes using XGBoost and SHAP

Abstract

Taxis play a crucial role in urban public transportation, but the traffic safety situation of taxi drivers is far from optimistic, especially considering the introduction of ride-hailing services into the taxi industry. This study conducted a comparative analysis of risk factors in crashes between traditional taxi drivers and ride-hailing taxi drivers in China, including their demographic characteristics, working conditions, and risky driving behaviors. The data was collected from 2,039 traditional taxi drivers and 2,182 ride-hailing taxi drivers via self-reported questionnaires. Four XGBoost models were established, taking into account different types of taxi drivers and crash types. All models showed acceptable performance, and SHAP explainer was used to analyze the model results. The results showed that for both taxi drivers, risk factors related to risky driving behaviors are more important in predicting property damage (PD) crashes, while risk factors related to working conditions are more important in predicting person injury (PI) crashes. However, the relative importance of each risk factor varied depending on the type of crashes and the type of taxi drivers involved. Furthermore, the results also validated certain interactions among the risk factors, indicating that the combination of certain factors generated a greater impact on crashes compared to individual factors alone. These findings can provide valuable insights for formulating appropriate measures to enhance road safety for taxi driver.

Keywords:

• Traditional taxi drivers
• ride-hailing taxi drivers
• taxi-related crashes
• comparative analysis
• safety implications

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This work was supported by the National Natural Science Foundation of China under Grant No. 52302443 and the Specialized Research Project of Shaanxi Provincial Department of Education under Grant No. 23JK0126.