Cross-State Traffic Accident Risk Prediction Using Self-Supervised Representation Learning on the US-Accidents Dataset
The Saudi Journal of Applied Sciences and Technology · 20.04.2026 · 2 (1)
DOI: 10.63908/hmw5tk26 ↗Kurz gefasst
Fachartikel in der Zeitschrift The Saudi Journal of Applied Sciences and Technology, online erschienen am 20.04.2026. Verfasst von Hani Alnami. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Traffic accident risk prediction is critical for intelligent transportation systems; however, models trained in one region often degrade when applied to other states due to heterogeneous traffic patterns, weather and infrastructure. This study evaluates cross-state accident risk prediction on the US-Accidents dataset (2016-2023; 7.7M records) under four transfer protocols: state-to-state, state-to-region, region-to-state, and region-to-region. A self-supervised learning with fine-tuning (SSL+FT) framework is proposed, where an encoder is pretrained on unlabeled data using a hybrid contrastive and masked reconstruction objective, then fine-tuned for binary risk classification. The approach is compared with traditional machine learning models (logistic regression, random forest, XGBoost) and deep learning models (multilayer perceptron (MLP), long short-term memory (LSTM), bidirectional LSTM (BiLSTM)). Results confirm significant performance degradation under cross-state distribution shift. Random forest and XGBoost achieve the best overall performance, with F1-scores up to 78% and 77%, respectively, while deep learning models show competitive but less stable performance across regions. In contrast, the proposed SSL+FT approach demonstrates consistent robustness, achieving up to 74% F1-score and 90% recall in region-to-region transfer. These findings indicate that self-supervised representation learning provides a scalable and generalizable solution for cross-state accident risk prediction under distribution shift.
Zitieren
APA
Alnami, H. (2026). Cross-State Traffic Accident Risk Prediction Using Self-Supervised Representation Learning on the US-Accidents Dataset. The Saudi Journal of Applied Sciences and Technology, 2(1). https://doi.org/10.63908/hmw5tk26
Vancouver
Alnami H. Cross-State Traffic Accident Risk Prediction Using Self-Supervised Representation Learning on the US-Accidents Dataset. The Saudi Journal of Applied Sciences and Technology. 2026;2(1). doi:10.63908/hmw5tk26
BibTeX
@article{Alnami2026,
author = {Alnami, Hani},
title = {Cross-State Traffic Accident Risk Prediction Using Self-Supervised Representation Learning on the US-Accidents Dataset},
journal = {The Saudi Journal of Applied Sciences and Technology},
year = {2026},
volume = {2},
number = {1},
doi = {10.63908/hmw5tk26}
}
EUVIMED