An Interpretable Heterogeneous Stacking Ensemble Model for Urban Traffic Crash Severity Classification
IET Intelligent Transport Systems · 2026-01 · 20 (1)
DOI: 10.1049/itr2.70382 ↗Kurz gefasst
Fachartikel in der Zeitschrift IET Intelligent Transport Systems, online erschienen am 17.11.2026. Verfasst von Mengmeng Yang und 2 weiteren Autor:innen. Die Arbeit nennt 34 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Jiangsu Provincial Department of Education.
Abstract (Originalsprache)
To address class imbalance, feature redundancy, limited information utilisation by individual models and insufficient interpretability in traffic crash severity prediction, this study develops a two‐layer heterogeneous stacking framework integrating sample resampling, grouped feature selection and probability fusion using crash data from Tempe, Arizona, USA, for 2021–2023. SMOTENC augments minority‐class samples within the training data, while recursive feature elimination with cross‐validation (RFECV), grouped by original variable, selects model‐specific feature subsets for random forest (RF) and XGBoost. Bayesian optimisation determines hyperparameters. Stratified five‐fold cross‐validation generates out‐of‐fold class probabilities from both base learners for integration by a CatBoost meta‐learner. Performance is evaluated using accuracy, weighted F1 and macro‐averaged AUC, supplemented by balanced accuracy and MCC. Ablation comparisons assess module contributions, while SHAP and LIME provide global and local explanations of model predictions, respectively. Under the current evaluation scenario, the framework achieves an accuracy of 82.10%, a weighted F1 of 82.25% and a macro‐averaged AUC of 0.821, outperforming all evaluated baselines on these metrics. Accuracy and weighted F1 exceed those of the best individual model for each metric by 3.25 and 3.45 percentage points, respectively. Ablation comparisons support the contributions of SMOTENC, grouped RFECV and both base learners. SHAP identifies collision manner, participant type and participant actions as key predictors, with contribution rankings and directions varying across severity classes and samples. LIME further reveals differences in base‐learner responses to crash variables and the meta‐learner's use of class probabilities. By combining heterogeneous model fusion with explanations at both decision layers, the framework provides a traceable basis for urban crash severity classification, risk stratification and traffic safety management.
Förderung
- Jiangsu Provincial Department of Education
Zitieren
APA
Yang, M., Sun, C., & Liu, Y. (2026). An Interpretable Heterogeneous Stacking Ensemble Model for Urban Traffic Crash Severity Classification. IET Intelligent Transport Systems, 20(1). https://doi.org/10.1049/itr2.70382
Vancouver
Yang M, Sun C, Liu Y. An Interpretable Heterogeneous Stacking Ensemble Model for Urban Traffic Crash Severity Classification. IET Intelligent Transport Systems. 2026;20(1). doi:10.1049/itr2.70382
BibTeX
@article{Yang2026,
author = {Yang, Mengmeng and Sun, Chao and Liu, Ying},
title = {An Interpretable Heterogeneous Stacking Ensemble Model for Urban Traffic Crash Severity Classification},
journal = {IET Intelligent Transport Systems},
year = {2026},
volume = {20},
number = {1},
doi = {10.1049/itr2.70382}
}
EUVIMED