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

Area of Science:

• Traffic Safety
• Machine Learning
• Causal Inference

Background:

• Advanced machine learning techniques lack interpretability in decision-making.
• Data-driven causal discovery methods often fail to provide clear explanations.
• Need for transparent models in complex domains like traffic safety.

Purpose of the Study:

• Propose a semi-data-driven Bayesian network structure learning algorithm, the Expert Knowledge Constraint-based (EKC) algorithm.
• Enhance the interpretability of causal relationship discovery in traffic crash analysis.
• Integrate expert knowledge with data for more transparent and accurate causal models.

Main Methods:

• Developed the Expert Knowledge Constraint-based (EKC) algorithm, integrating expert knowledge with conditional independence tests.
• Applied the EKC algorithm to highway crash data from the HuNing Highway (2022).
• Utilized Bayesian estimation and variable elimination algorithms for effect estimation and scenario ranking.

Main Results:

• Date-related variables were found not to directly influence crashes.
• Unfavorable temperatures, medium traffic volumes, and snowy weather correlate with increased crash probabilities.
• Highest crash probability identified under specific conditions: medium traffic, cold temperatures, winter, cloudy weather, morning, and weekdays.

Conclusions:

• The EKC algorithm offers superior interpretability compared to Hill Climbing, Chow-Liu Trees, and logistic models, while maintaining strong fitting scores.
• Established a framework for model interpretability in traffic crash analytics, encompassing causality, trust, heterogeneity, transferability, and stability.
• The EKC algorithm provides a more transparent and interpretable approach to understanding and mitigating highway crash risks.