Application of explainable machine learning for real-time safety analysis toward a connected vehicle environment

Chen Yuan; Ye Li; Helai Huang; Shiqi Wang; Zhenhao Sun; Honggang Wang

doi:10.1016/j.aap.2022.106681

Application of explainable machine learning for real-time safety analysis toward a connected vehicle environment

Accid Anal Prev. 2022 Jun:171:106681. doi: 10.1016/j.aap.2022.106681. Epub 2022 Apr 22.

Authors

Chen Yuan¹, Ye Li², Helai Huang³, Shiqi Wang⁴, Zhenhao Sun⁵, Honggang Wang⁶

Affiliations

¹ School of Traffic and Transportation Engineering, Central South University, Changsha, Hunan 410075, China. Electronic address: yuanchen@csu.edu.cn.
² School of Traffic and Transportation Engineering, Central South University, Changsha, Hunan 410075, China. Electronic address: yelicsu@csu.edu.cn.
³ School of Traffic and Transportation Engineering, Central South University, Changsha, Hunan 410075, China. Electronic address: huanghelai@csu.edu.cn.
⁴ Department of Computer Science, City University of Hong Kong, Kowloon, Hong Kong, China. Electronic address: shiqwang@cityu.edu.hk.
⁵ Department of Computer Science, City University of Hong Kong, Kowloon, Hong Kong, China. Electronic address: zhsun.cs@my.cityu.edu.hk.
⁶ School of Traffic and Transportation Engineering, Central South University, Changsha, Hunan 410075, China. Electronic address: wanghonggang@caeri.com.cn.

PMID: 35468530
DOI: 10.1016/j.aap.2022.106681

Abstract

Due to the difficulty of obtaining traffic flow data and conflicts simultaneously, conflict-based analysis using macroscopic traffic features is much less studied. This research aims to analyze real-time safety by a disaggregate study and explore the benefit of the connected vehicle (CV) for real-time safety evaluation. To avoid the endogeneity problem regarding conflicts and traffic features in regression models, machine learning is employed to obtain a reliable and practical real-time safety model. The results show that the Random Forest outperforms eXtreme Gradient Boosting, Support Vector Machine and Adaptive Boosting models, achieving the best performance with the highest AUC of 0.827. For a deep understanding of conflict mechanisms, the explainable machine learning method SHAP (SHapley Additive exPlanation) is introduced to improve the model interpretability providing insights into the impacts of traffic flow features. Lane difference regarding average speed is found to have the most significant impacts on real-time safety. Speed variation, the proportion of trucks and traffic volume are associated with conflict occurrence. Further analysis highlights that the impacts of traffic features are heterogeneous and there may exist specific patterns of paired features affecting real-time safety. Encouragingly, SHAP appears to be able to complement the traditional model with random components in terms of revealing heterogeneity. The explainable machine learning can also provide a solid basis for discretizing continuous variables while previous studies perform discretization mainly based on prior knowledge and experience. The experimental result regarding CV Market Penetration Rate (CV-MPR) demonstrates that the model performance is gradually elevated with the increase of penetration rate. The initial stage of the CV market (20%, 40% CV-MPR) yields the most significant gains in real-time safety evaluation. These findings can be used beneficially in active traffic management.

Keywords: Connected vehicles; Machine learning; Market penetration rate; Real-time safety; SHAP.

MeSH terms

Accidents, Traffic* / prevention & control
Humans
Machine Learning
Motor Vehicles*
Safety
Support Vector Machine