Causal relationship discovery for highway crash analysis using semi-data-driven Bayesian network.

Abstract

With the widespread application of advanced machine learning techniques, researchers need a more transparent decision-making process. The data-driven causal relationship discovery techniques often lack interpretability. Therefore, a semi-data-driven Bayesian network structure learning algorithm, the Expert Knowledge Constraint-based (EKC) algorithm, is proposed. By integrating expert knowledge with conditional independence tests, the EKC algorithm constructs a causal Bayesian network with a high level of interpretability. The algorithm was applied to a highway safety scene using crash data collected in 2022 from the HuNing Highway in China. The effects of the Bayesian network on variables were estimated using the Bayesian estimation algorithm, and the most dangerous scenarios were ranked using the variable elimination algorithm. Key findings include: (1) date-related variables do not directly affect crashes; (2) unfavorable temperatures, medium-level traffic volumes, and snowy weather conditions are associated with higher crash probabilities; and (3) the highest crash probability occurs under medium traffic volume, cold temperatures, winter season, cloudy weather, morning hours, and weekdays. The EKC algorithm was compared with the Hill Climbing algorithm, Chow-Liu Trees algorithm, and logistic model, demonstrating significant improvements in interpretability while maintaining good fitting scores. Furthermore, the definition framework of model interpretability in traffic crash analytics was discussed, including causality, trust, heterogeneity, transferability, and stability.