Lightweight and Interpretable Transformer-Based Models for Predicting Driving Comfort in Autonomous Vehicles

Accurate prediction of driving comfort is important for improving both user experience and operational safety in autonomous vehicles. This study examines lightweight and interpretable Transformer-based models, including Multi-Head Attention, FT-Transformer, and SAINT, to evaluate their ability to predict driving comfort efficiency and reliably. A case study on publicly available real-world road datasets is conducted, and the proposed models are compared with a Transformer model used in an earlier study. Computational complexity is analyzed using Big-O notation. Results demonstrate that the Transformer-based models evaluated in this work achieve improved computational efficiency and maintain competitive predictive performance. After feature selection and a reduction in the number of road-related input features, the models maintain similar levels of accuracy. These findings indicate that accurate comfort prediction can be achieved with a reduced feature set and efficient model designs. The proposed framework thus provides a possible practical solution for real-time integration into autonomous vehicle systems that supports both safety and ride comfort.