Identifying the latent relationships between factors associated with traffic crashes through graphical models

Area of Science:

• Traffic Safety
• Data Science
• Machine Learning

Background:

• Traditional traffic safety models focus on direct relationships between crash outcomes and predictor variables.
• Existing methods often overlook latent relationships between crash factors and lack tools for informed variable selection, especially with limited data.
• Understanding complex interactions is crucial for effective crash prevention.

Purpose of the Study:

• To propose and apply graphical models for analyzing fatal and incapacitating injury pedestrian crashes.
• To disclose the relationship topologies of explanatory variables contributing to pedestrian crashes.
• To address limitations in current traffic safety modeling regarding latent relationships and variable selection.

Main Methods:

• Utilized graphical models, including Markov random field (MRF) modeling, Bayesian network modeling, and a graphical XGBoost approach.
• Applied these models to identify the structure and essential factors involved in pedestrian crashes.
• Focused on uncovering latent relationships between variables influencing crash outcomes.

Main Results:

• The study demonstrates the potential of graph learning models in traffic safety research.
• Identified complex variable structures and essential factors contributing to pedestrian crashes.
• Revealed latent relationships that are often missed by traditional modeling techniques.

Conclusions:

• Graphical models offer a powerful approach to understand the mechanisms behind crash occurrences.
• These methods can assist in developing more accurate and reliable prevention measures by identifying critical variables.
• The application of graph learning in traffic safety provides insights akin to a pathological examination for understanding crash causality.