A hybrid approach for modeling bicycle crash frequencies: Integrating random forest based SHAP model with random...

Area of Science:

• Transportation Safety
• Data Science
• Statistical Modeling

Background:

• Bicycle crash frequency analysis requires models that capture complex relationships and unobserved factors.
• Existing methods often struggle with nonlinearities and heterogeneity in crash data.

Purpose of the Study:

• To develop and validate a novel hybrid framework integrating Random Forest-based SHapley Additive exPlanations (RF-SHAP) with a random parameter negative binomial regression (RPNB) model.
• To enhance the explanation of complex, nonlinear relationships in bicycle crash frequency data while accounting for unobserved heterogeneity.

Main Methods:

• Comparative analysis of four machine learning algorithms (RF, SVM, GBM, XGBoost) for variable importance.
• Integration of the best-performing Random Forest algorithm with SHAP for interpretable variable impact assessment (RF-SHAP).
• Combination of the RF-SHAP method with the RPNB model to analyze individual-specific variations influencing crash predictions.

Main Results:

• The proposed RF-SHAP and RPNB framework demonstrated superior prediction accuracy for bicycle crash frequency.
• Improved interpretability of risk factors influencing bicycle crashes was achieved through SHAP values and RPNB causal insights.
• The model showed consistent Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) values, indicating reliable explanatory power.
• Significant improvements were observed in Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE).

Conclusions:

• The hybrid framework effectively combines the explanatory power of statistical models with the forecasting capabilities of data-driven models.
• The interpretable SHAP values and RPNB causal insights offer actionable information for policymakers to develop targeted bicycle safety interventions.
• This approach provides a robust tool for understanding and mitigating bicycle-related traffic incidents.