Adaptive ensemble decision integration for indicator of resource security: methodology and statistical validation of stability

• Дата: 2026
• Автори: Ilyina, O.P., Skybyk, S.Ya.
• Формат: Стаття
• Мова: Українська
• Опубліковано: PROBLEMS IN PROGRAMMING 2026
• Теми: machine-learning classification; ensemble training; adaptive quality metric; class imbalance; soft voting; statistical validation; strategic decision support; UDC 004.855; 519.86
• Онлайн доступ: https://pp.isofts.kiev.ua/index.php/ojs1/article/view/879
• Назва журналу: Problems in programming

Ensuring effective decision support in complex distributed organizational systems (especially in national security and defense planning) requires reliable classification methods capable of rapid diagnosis of resource states and risks to strategic interests. The effectiveness of a resource security indicator (RSI) built on machine learning methods critically depends on the stability and reliability of integrated predictions under conditions typical of this domain: significant class imbalance (where missing a negative state is critical), limited data volume, log normal feature distribution with "long tails", and noise components that reduce the stability of individual classifiers. To address these challenges, an adaptive ensemble integration mechanism (RSI) was developed, implementing weighted soft voting of models (NB, SVM, RF, kNN, LR) with unified probability calibration. The central element is a composite dynamic quality metric (KQ), which combines 1 (prioritizing the minority class), , and , adapting their weights based on correlation. Trust coefficients (KDR) are integrated to adjust the influence of models depending on their vulnerability to data properties. Algorithm validation was performed on synthetic data simulating log-normal distribution and lag effects of real-world conditions. A large-scale experiment (250 runs, paired design) confirmed high statistical significance ( 0.001 by Wilcoxon test) of RSI superiority over the best single classifier (Random Forest) across all metrics (Δ, Δ1, Δ). The effect size (Cohen's ≥ 1.41) indicates large practical value. The results demonstrate that adaptive integration ensures stability and reliability of risk diagnosis, critically necessary for security applications.Problems in programming 2025; 4: 88-101