EUVIMEDEUVIMED entdecken
Crossref

DriveStyleFormer: A Hybrid TCN-BiGRU-Transformer Network for Leakage-Aware Driving-Style Classification

Waiss Ali Aden, Tahsin Ullah, Kamran Khan

Journal of Machine Learning Advances · 2026

DOI: 10.66879/jmla.2026.54 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Machine Learning Advances, 2026. Verfasst von Waiss Ali Aden und 2 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Accurate driving-style recognition is important for advanced driver-assistance systems, fleet risk management, personalized vehicle control, and intelligent transportation. Existing approaches often treat driving records as independent observations or rely on a single temporal architecture, limiting their ability to jointly capture local maneuvers, sequential behavioral evolution, and longer-range dependencies. This study proposes DriveStyleFormer, a hybrid temporal deep-learning network integrating dilated temporal convolutional blocks, a two-layer bidirectional gated recurrent unit, a Transformer encoder, gated feature fusion, and temporal attention pooling. The analysis used 162,566 observations from 12,994 trips and incorporated ego-vehicle, preceding-vehicle, traffic, weather, road, and lighting information. Temporal and interaction-related features were engineered to construct 39 transformed inputs. To reduce trip-level information leakage, trip identifiers were partitioned into independent training, validation, and held-out test sets, and each prediction was based on a causal window of up to 16 current-and-past observations. DriveStyleFormer was compared with multilayer perceptron and standalone GRU baselines under identical data partitions and optimization settings. On the held-out test set, DriveStyleFormer achieved a macro-F1 score of 0.7325 and a macro one-vs-rest ROC-AUC of 0.9045, outperforming both baseline models. Class-wise analysis showed the strongest performance for Aggressive driving, while greater confusion was observed between the Normal and Vague classes. These findings indicate that combining complementary temporal representations with leakage-aware evaluation can improve driving-style classification. Further repeated grouped evaluation, broader baseline comparisons, ablation analysis, and external validation are required to establish broader generalizability.

Zitieren

APA

Aden, W. A., Ullah, T., & Khan, K. (2026). DriveStyleFormer: A Hybrid TCN-BiGRU-Transformer Network for Leakage-Aware Driving-Style Classification. Journal of Machine Learning Advances. https://doi.org/10.66879/jmla.2026.54

Vancouver

Aden WA, Ullah T, Khan K. DriveStyleFormer: A Hybrid TCN-BiGRU-Transformer Network for Leakage-Aware Driving-Style Classification. Journal of Machine Learning Advances. 2026. doi:10.66879/jmla.2026.54

BibTeX
@article{Aden2026,
  author = {Aden, Waiss Ali and Ullah, Tahsin and Khan, Kamran},
  title = {DriveStyleFormer: A Hybrid TCN-BiGRU-Transformer Network for Leakage-Aware Driving-Style Classification},
  journal = {Journal of Machine Learning Advances},
  year = {2026},
  doi = {10.66879/jmla.2026.54}
}
Kostenlos starten →