Predicting Future Driving Risk of Crash-Involved Drivers Based on a Systematic Machine Learning Framework.

Area of Science:

• Traffic Safety
• Machine Learning
• Data Science

Background:

• Predicting future driving risk is crucial for targeted safety interventions.
• Existing methods often lack the sophistication to handle complex driving behavior data.

Purpose of the Study:

• To develop a machine learning framework for predicting future driving risk in crash-involved drivers.
• To define driving risk, identify key risk factors, and build a reliable predictive model.

Main Methods:

• A machine learning framework was applied to seven-year crash/violation records.
• Drivers were categorized into high-risk (HR) and low-risk (LR) groups based on five scenarios.
• Four tree-based ensemble methods (RF, Adaboost, GBDT, XGboost) were evaluated using a one-year rolling time window.

Main Results:

• A multi-dimensional risk definition including crash recurrence, severity, and fault commitment was optimal.
• Gradient Boosting Decision Tree (GBDT) demonstrated the best performance, achieving an average precision (AP) of 0.68.
• Seven of the top nine predictive features were linked to risky driving behaviors with non-linear relationships.

Conclusions:

• The proposed machine learning approach effectively predicts future driving risk.
• GBDT is a suitable model for this task, and risky driving behaviors are key predictors.
• Accurate risk prediction enables more effective traffic safety interventions.