A Meta-Ensemble Framework for Road Accident Prevention using Transformer and Traditional Machine Learning Models

View at publisher Summarise with VIRAT
Authors: John Nathanael J, Vivek Sharma S, Abdul Hashim S, Maheswari S
Cited by: 0
References: 14

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Accurately predicting road accidents is vital for enhancing traffic safety and reducing fatalities. Conventional machine learning techniques often face challenges in handling the heterogeneous characteristics of structured accident datasets that contain both categorical and numerical variables. Early prediction of accident risk can enable intelligent transportation systems to mitigate hazards and improve traffic safety. Our proposal is a hybrid predictive framework that integrates a Feature Tokenizer Transformer (FT-Transformer) with classical ensemble learners, namely Random Forest and XGBoost, combined through a metalearner (Logistic Regression). Using an upsampled dataset of 1200 accident records from kaggle, the proposed hybrid model achieves an accuracy with value of 0.80 and$\mathbf{F 1}$-score with value of 0.79, significantly outperforming baseline models such as Logistic Regression (Accuracy$=0.58$, F1$=0.61$) and shallow Decision Trees (Accuracy$\boldsymbol{=} \mathbf{0. 6 7}, \mathbf{F 1} \boldsymbol{=} \mathbf{0. 6 8}$). Our results demonstrate the capabilities of transformer-based architectures to model complex feature interaction in tabular data, while tree ensembles provide robustness and interpretability. The proposed framework offers practical potential for real-time accident risk alerts in intelligent transportation systems.

Fields of Study & Subjects

Traffic and Road Safety IoT and GPS-based Vehicle Safety Systems Traffic Prediction and Management Techniques

Publication details

Journal / source
International Conference on Intelligent Sustainable Systems (ICISS)
Pages
974–981
Publication date
4 March 2026
Language
EN