Assessing the impact of car-following driving style on traffic conflict risk using asymmetric behavior model and explainable machine learning.

Abstract

To deepen the understanding of the impact of car-following driving style (CFDS) on traffic conflict risk and address the lack of clear CFDS evaluation metrics, this study proposes an improved CFDS metric based on the Asymmetric Behavior (AB) theory. Interpretable machine learning models were utilized for regression analysis to examine the relationship between CFDS and conflict risk. The generalized AB model calculates the difference between vehicle trajectories and the Newell trajectory, constructing the driving style evaluation metric, which quantifies driver aggressiveness in a manner that is both computationally straightforward and easily interpretable. High-precision vehicle trajectory data were collected using radar-camera integrated devices, enabling the use of various interpretable machine learning methods to model and analyze the impact of driving style on conflict risk. The results demonstrate that the proposed car-following driving style evaluation metric consistently shows the highest importance across multiple datasets with different risk levels and sampling windows, indicating a strong correlation with conflict risk. Interpretations using Shapley Additive Explanations reveal a nuanced, yet mostly monotonic impact pattern of driving style across high, medium, and low-risk scenarios, with more aggressive drivers being more prone to high-risk situations. Furthermore, Partial Dependence Plot analysis reveals a complex, saddle-shaped risk curve related to driving style and its interactions, highlighting that aggressive and "pseudo-timid" drivers exhibit higher risks in specific contexts. In summary, this research constructs clear and interpretable CFDS evaluation metrics, validated through case analysis for their rationality and effectiveness, thereby providing new theoretical support for traffic risk prediction and intervention.