Class Imbalance Strategies, Calibration, and Explanation Stability in Tree-Based Intrusion Detection for IoT Networks
Journal of Expert Systems and Sustainable Development · 2026
DOI: 10.65069/jessd21202623 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Expert Systems and Sustainable Development, 2026. Verfasst von Abdulkerim Oğuzhan Alkan.
Abstract (Originalsprache)
Network intrusion detection for the Internet of Things is based on traffic dominated by a few attack families, while security-relevant classes are rare. The credibility of published results depends on evaluation choices that are seldom examined together. Treatments for class imbalance are typically validated within a single learner family on a fixed training set, leaving the interaction between strategy and learner unclear and exposing reported gains to seed dependence. This paper compares five class imbalance strategies using two tree-based learners across the binary, category, and attack-type levels of the CICIoT2023 benchmark. The training pools are sampled from the full population using ten independent seeds, and the configurations are compared using matched protocols and nonparametric tests. In the controlled comparison, class weighting is the strongest single mechanism for gradient boosting, significantly outperforming every alternative. The balance between weighting and synthetic oversampling varies across learner families. Probability calibration separates the two model families by more than an order of magnitude in expected calibration error, while TreeSHAP rankings remain robust to resampling. The findings support a learner-dependent view of imbalance treatment and position calibration and explanation stability as evaluation axes in their own right, with implications for threshold-based alerting and the forensic use of feature attributions.
Zitieren
APA
Alkan, A. O. (2026). Class Imbalance Strategies, Calibration, and Explanation Stability in Tree-Based Intrusion Detection for IoT Networks. Journal of Expert Systems and Sustainable Development. https://doi.org/10.65069/jessd21202623
Vancouver
Alkan AO. Class Imbalance Strategies, Calibration, and Explanation Stability in Tree-Based Intrusion Detection for IoT Networks. Journal of Expert Systems and Sustainable Development. 2026. doi:10.65069/jessd21202623
BibTeX
@article{Alkan2026,
author = {Alkan, Abdulkerim Oğuzhan},
title = {Class Imbalance Strategies, Calibration, and Explanation Stability in Tree-Based Intrusion Detection for IoT Networks},
journal = {Journal of Expert Systems and Sustainable Development},
year = {2026},
doi = {10.65069/jessd21202623}
}
EUVIMED