Investigating the contributors to hit-and-run crashes using gradient boosting decision trees

Area of Science:

• Traffic Safety
• Data Science
• Machine Learning

Background:

• Hit-and-run crashes pose a significant public safety concern.
• Predicting perpetrator escape behavior is crucial for prevention and investigation.
• Existing models may not fully capture the complexity of hit-and-run incidents.

Purpose of the Study:

• To develop and evaluate a Gradient Boosting Decision Tree (GBDT) model for predicting perpetrator escape behavior in hit-and-run crashes.
• To compare the performance of GBDT against other classification methods using the U.S. Crash Report Sampling System (CRSS) dataset.
• To identify key contributing factors and interactions influencing escape behavior.

Main Methods:

• Utilized the U.S. Crash Report Sampling System (CRSS) dataset.
• Developed a classification prediction model using the Gradient Boosting Decision Tree (GBDT) algorithm.
• Compared GBDT with Classification and Regression Tree (CART), Random Forest, and Logistic Regression.

Main Results:

• GBDT achieved superior performance with the lowest negative log-likelihood (0.282) and misclassification rate (0.096), and the highest AUC (0.803).
• GBDT demonstrated high computational efficiency (LIFT value of 4.087).
• Identified crash type and relation to trafficway as significant factors, uncovering previously unhighlighted information.

Conclusions:

• The GBDT model is a more accurate and efficient tool for predicting hit-and-run crashes compared to traditional methods.
• The model's ability to identify hidden factors and variable interactions offers valuable insights for accident analysis.
• Findings have practical implications for hit-and-run incident prevention, traffic safety analysis, and engineering applications.