A two-country naturalistic driving study conducted across Belgium and the United Kingdom found that data augmentation using Conditional Generative Adversarial Networks maintained minority-class classification performance for high-risk driving behaviors but did not consistently surpass unaugmented baselines under rigorous evaluation. Evaluating driver risk across real-world trips remains critical because traffic collisions claim approximately 1.19 million lives annually worldwide, with human error contributing to over 90 percent of recorded vehicle crashes through actions such as excessive speeding, harsh acceleration, and sudden braking maneuvers. While intelligent transportation systems and advanced vehicle telematics allow researchers to analyze real-world driving behaviors, operational datasets suffer from severe class imbalance because catastrophic events and near-collisions occur infrequently compared to normal driving routines. To investigate whether synthetic sample generation could resolve this training constraint, researchers evaluated a generative framework combining conditional synthetic sample generation with three supervised classifiers, tracking detection rates alongside model interpretability across real-world telematics gathered over four months.

Naturalistic driving studies provide authentic insight into actual motorist behaviors in regular road environments, contrasting with controlled driving simulator studies that lack naturalistic operational complexity. Data for this investigation originated from the European i-DREAMS project, which monitored driving parameters using On-Board Diagnostics devices installed in passenger vehicles across Belgium and the United Kingdom. Data collection tracked 43 participants from Belgium and 26 participants from the United Kingdom across four distinct experimental phases, continuously recording vehicle telematics while examining the influence of real-time warning systems and post-drive feedback. Raw driving records were segmented into discrete 30-second time slices, with statistical aggregates calculated for vehicle operational metrics such as mean travel speed, trip distance, and braking intensity. Missing telematics entries recorded as negative values were systematically cleaned, and numerical parameters were normalized using Min-Max scaling to ensure uniform feature representations across experimental modeling pipelines.

To categorize driver risk objectively, the i-DREAMS operational framework originally established a Safety Tolerance Zone structured into three risk tiers labeled Normal, Dangerous, and Avoidable Accident. These designations were determined through predefined intervention thresholds tracking real-time speeding and headway distance deviations, encoded via hierarchical binary flags from minus one to three. Because intermediate Dangerous events represented an ambiguous middle tier with low sample volumes that risked introducing boundary noise, researchers restructured the classification task into a focused binary problem contrasting Normal driving against Avoidable Accident events. Normal intervals corresponded to intervention levels of minus one, zero, or one, whereas Avoidable Accident intervals reflected the highest critical risk tier at level three. The resulting training datasets reflected severe real-world class distributions: Normal intervals comprised more than 80 percent of records in Belgium and approximately 94 percent in the United Kingdom, leaving Avoidable Accidents as a rare minority.

To generate realistic synthetic samples of uncommon risk events, the research team deployed a Conditional Generative Adversarial Network trained on the full imbalanced training dataset with class labels supplied as conditioning inputs to both generator and discriminator networks. The generator network utilized fully connected layers with batch normalization and Leaky ReLU activation functions, mapping random noise vectors and target risk labels into synthetic driving observations. The discriminator network incorporated convolutional layers applied across the temporal dimension of the 30-second driving windows and was optimized using a Wasserstein loss function with a gradient penalty parameter of ten. This adversarial configuration was specifically selected over standard cross-entropy formulations to provide stable optimization gradients and mitigate mode collapse. During synthetic data generation, the trained generator was conditioned on the Avoidable Accident label to synthesize minority-class examples, which were then merged with real driving observations to assemble balanced training datasets.

To rigorously assess classification performance before and after synthetic augmentation, the researchers examined three distinct machine learning architectures capable of processing multivariate vehicle telematics. The first architecture comprised a hybrid XGBoost and Random Forest ensemble, in which Random Forest bagging initially filtered noise and provided feature importance rankings while Extreme Gradient Boosting optimized the final classification decision boundary. The second architecture utilized a Recurrent Neural Network paired with an AdaBoost algorithm, incorporating Long Short-Term Memory units to capture temporal dependencies across sequential sensor observations. The third model adapted the convolutional discriminator component of the Conditional Generative Adversarial Network directly as a standalone binary classifier tasked with distinguishing Normal from Avoidable Accident intervals.

To avoid optimistic evaluation artifacts and prevent data leakage, the study enforced a strict participant-level stratified holdout protocol. Participants were allocated entirely to either training and validation sets or independent test sets, ensuring no participant contributed driving trips to both subsets. Across the cohort, 80 percent of participants were designated for model development—split internally into 87.5 percent training and 12.5 percent validation—while 20 percent were reserved exclusively for final evaluation. For the Belgian cohort, 34 drivers were assigned to training and validation and nine drivers to testing; for the British cohort, 21 drivers supported development and five drivers served as held-out test subjects. All generative network training and synthetic sample generation occurred exclusively on data from development participants, and no synthetic instances were ever introduced into validation or test sets, ensuring model evaluation occurred solely against authentic driving data.

Classifier performance was tracked using precision, recall, accuracy, false alarm rates, and minority-class F1-scores, contrasting conditional generative augmentation against an unaugmented baseline and four standard rebalancing methods: class weighting, random oversampling, SMOTE, and ADASYN. When trained on unaugmented data, all classifiers demonstrated high overall accuracy alongside markedly lower minority-class F1-scores, demonstrating that standard accuracy metrics mask detection failures for rare safety-critical events. In the Belgian dataset, the unaugmented baseline achieved the strongest minority-class F1-scores for XGBoost-RF at 0.646, RNN-AdaBoost at 0.638, and the GAN classifier at 0.638. Following conditional generative augmentation, XGBoost-RF achieved an F1-score of 0.640 and RNN-AdaBoost maintained 0.638, closely matching baseline levels without improvement, whereas the standalone GAN classifier dropped to 0.417. In the United Kingdom dataset, where class imbalance was severe, XGBoost-RF matched its baseline score of 0.313, while RNN-AdaBoost declined from 0.247 to 0.190, and the GAN classifier rose slightly from 0.141 to 0.155.

Comparative evaluations against standard baseline resampling strategies further demonstrated that synthetic augmentation does not consistently outperform simpler imbalance-handling techniques when applied to low-dimensional driving data. Across both national datasets, ADASYN consistently yielded the lowest performance metrics among all evaluated classifiers, while class weighting produced competitive or superior outcomes relative to complex oversampling. In the United Kingdom dataset, class weighting achieved the highest minority-class F1-scores for both RNN-AdaBoost at 0.254 and the GAN classifier at 0.168, slightly outperforming the unaugmented baseline. Analysis of the generative model indicated that the generator suffered from mode collapse, evidenced by near-zero intra-class pairwise distances among generated samples. This low sample diversity limited the representativeness of synthetic Avoidable Accident instances, introducing uninformative noise into downstream decision boundaries rather than reinforcing genuine feature variance.

To evaluate how synthetic training data altered internal model logic, the authors implemented SHAP feature attribution to quantify the contribution of individual variables toward model predictions. In the unaugmented Belgian models, speed metrics including engine speed and mean GPS speed dominated predictions of Avoidable Accidents, followed by total distance, while harsh braking frequency had minimal influence. In the unaugmented British models, mean vehicle speed similarly emerged as the primary determinant, with trip distance and GPS speed providing secondary contributions, while inter-beat intervals representing heart rate variability contributed minimally. Following conditional generative augmentation, SHAP distributions shifted substantially: speed dominance declined, and total distance in Belgium and physiological heart rate intervals in the United Kingdom became the most influential predictive features. Although elevated physiological stress aligns conceptually with driver agitation, the authors emphasized that attribution shifts may reflect synthetic data artifacts rather than authentic behavioral insights.

From an operational deployment perspective, inference latency benchmarks indicated that all three classification pipelines remain computationally feasible for near-real-time implementation, executing single 30-second window evaluations within 2 to 9 milliseconds on standard central processing units, with XGBoost-RF requiring 2 milliseconds and RNN-AdaBoost requiring 9 milliseconds. However, the authors highlighted critical operational constraints, noting that geographic scope was confined to two Western European countries sharing similar infrastructure and that physiological monitoring was available only within the British cohort. Furthermore, differing sensor configurations between countries prevented direct performance equivalence. The researchers concluded that deploying generative models in safety-critical vehicle assistance systems requires external validation on independent driver populations, recommending that future investigations examine diversity-sensitive generative models, specialized tabular synthesis architectures such as CTGAN or TabDDPM, and multi-modal inputs combining vehicle telematics with video feeds and gaze-tracking data.

Subject of Research: Machine learning and generative data augmentation for risky driving classification

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

Article References: Theodoraki, E. M., Koliou, P., & Yannis, G. (2026). Improving risky driving behavior classification via conditional GAN-augmented hybrid machine learning models: a two-country naturalistic driving study. European Transport Research Review, 18(1), Article 71. https://doi.org/10.1186/s12544-026-00827-z

Image Credits: AI Generated

DOI: 10.1186/s12544-026-00827-z

Keywords: Machine Learning, Generative Adversarial Networks, Road Safety, Naturalistic Driving, XGBoost, Recurrent Neural Networks, SHAP Analysis, Class Imbalance, Improving, risky, driving, behavior