Improving risky driving behavior classification via conditional GAN-augmented hybrid machine learning models: a two-country naturalistic driving study

Abstract

Pinpointing and classifying risky driving behaviors is crucial for improving road safety and developing preventive interventions. However, practically collected driving datasets are highly imbalanced, with extreme risks such as reckless maneuvers and preventable collisions. To overcome this limitation, the current study introduces a hybrid machine learning pipeline that combines Conditional Generative Adversarial Networks (cGANs) for creating realistic synthetic samples with robust classifiers, including an XGBoost-Random Forest ensemble, a Recurrent Neural Network boosted by AdaBoost, and dedicated GAN-driven models. To prevent information leakage, data augmentation was applied only to the training subset after dataset partitioning. The original i-DREAMS driving risk labels consisted of three categories (Normal, Dangerous, and Avoidable Accident). To address severe class imbalance and focus on safety-critical events, the Dangerous and Avoidable Accident categories were restructured into a binary classification task: Normal versus Avoidable Accident. Synthetic minority-class examples were generated with cGANs and combined with real training data. Under a rigorous participant-level holdout evaluation protocol, cGAN augmentation maintained minority-class F1-scores comparable to the no-augmentation baseline across both countries, though it did not consistently exceed it. SHAP analysis revealed that augmentation shifted feature importance distributions, reducing speed-metric dominance and increasing the relative importance of distance and physiological features, but whether these shifts reflect genuine generalisation improvements or synthetic data artefacts requires further validation. The results demonstrate how cGAN-based augmentation maintains minority-class detection performance under severe class imbalance while preserving interpretability through SHAP analysis, though further validation on independent driver populations is required before deployment in safety-critical contexts. Nevertheless, additional validation on independent driver populations is required to assess the generalizability of the proposed framework.