A glass-box approach to driving risk: Explainable Boosting Machines for interpreting transient behaviors in trajectory data

Rising traffic accidents demand interpretable risk models. Using GPS data from 10,350 taxis in Beijing (T-Drive, ~16M records), trips were segmented into 5.87M 300-second windows. After DBSCAN and K-Means (k = 2, silhouette = 0.62), 14.85% of windows were labeled risky and 85.15% safe. For each window, we extracted features capturing driving instability: speed volatility, abrupt heading changes, and sudden stops. We compared an Explainable Boosting Machine (EBM) against black-box models (XGBoost, Random Forest) with SHAP. The EBM achieved 99.65% accuracy, 99.66% precision, 99.65% recall, and 99.66% F1-score, matching XGBoost while offering greater interpretability. High speeds, speed fluctuations, and abrupt direction changes were the the strongest risk indicators. The model explicitly reveals how individual features contribute to the risk score per window. These findings show that a glass-box model can match black-box performance with interpretability, providing a practical basis for auditing risky driving from large‑scale trajectory data and supporting transparent warning systems and usage‑based insurance.