SHAP-based interpretation of holistic traffic risk causation patterns using high-resolution trajectory data and risk...

Area of Science:

• Traffic Engineering
• Road Safety Analysis
• Machine Learning Applications

Background:

• Assessing road traffic safety using high-resolution trajectory data is crucial.
• Existing methods focusing on single conflict types fail to capture holistic road segment risk.
• A comprehensive approach is needed to evaluate overall traffic risk and identify influencing factors.

Purpose of the Study:

• Develop a method for evaluating road segment-level traffic risk irrespective of specific conflict types.
• Analyze key factors influencing traffic risks and explore their underlying mechanisms.
• Enhance the interpretability of machine learning-based traffic risk models.

Main Methods:

• Proposed an integrated surrogate safety measure based on risk field theory for comprehensive conflict identification.
• Developed road segment-level holistic risk assessment models using logistic regression and five machine learning methods.
• Applied resampling techniques to address imbalanced sample categories in traffic conflict data.

Main Results:

• The random forest model achieved 80.6% prediction accuracy, 89.0% precision, and 86.8% recall for conflict cases.
• SHAP analysis revealed the contribution of traffic state variables and their interactions to traffic risk.
• Vehicle type distribution in traffic flow was identified as a significant factor influencing conflict occurrence.

Conclusions:

• The study provides accurate risk prediction for road segments, moving beyond single conflict type analysis.
• Enhanced interpretability of machine learning models bridges prediction with causal insight into traffic risks.
• Findings contribute to a deeper understanding of traffic dynamics and road safety mechanisms.