Research indexJournal article
ДИВЕРСИФІКОВАНІ АНСАМБЛІ КЛАСИФІКАТОРІВ ДЛЯ ВИЯВЛЕННЯ ДЕФЕКТІВ ПРОГРАМНОГО КОДУ ЗА СТАТИЧНИМИ МЕТРИКАМИ ЯКОСТІ
Hornostal, Chelak
Системи управління навігації та зв’язку Збірник наукових праць18 Sep 2026Open access
NobleIDark:/48914/w1/20261105/1E456075DOI10.26906/sunz.2026.3.150
ДИВЕРСИФІКОВАНІ АНСАМБЛІ КЛАСИФІКАТОРІВ ДЛЯ ВИЯВЛЕННЯ ДЕФЕКТІВ ПРОГРАМНОГО КОДУ ЗА СТАТИЧНИМИ МЕТРИКАМИ ЯКОСТІ
Hornostal, Chelak · Системи управління навігації та зв’язку Збірник наукових праць · 2026
NobleIDark:/48914/w1/20261105/1E456075DOI10.26906/sunz.2026.3.150
Abstract
Background The relevance of the study lies in the need to improve existing approaches to source code defect classification in order to provide timely warning of potential execution problems and possible intrusions. The object of the study is the process of automated detection of software code defects. The subject of the study is ensemble machine learning classification methods. The objective of the study is to improve the quality of ensemble classifiers in software code defect classification tasks. The research tasks are: to analyze the characteristics of the dataset; to construct a sequence of ensemble models, from single classifiers to a diversified ensemble, for the classification of potential defects in software modules; and to confirm the statistical significance of the quality gain achieved at each stage. The methods used include data preprocessing techniques, feature selection methods, base machine learning methods, ensemble methods, diversification techniques, and statistical evaluation methods. Results The study experimentally confirmed a monotonic increase in classification quality as the ensemble architecture was progressively complexified – from the best-performing base classifier, Gradient Boosting (CV F1 = 0.798), to the proposed extended diversified stacking method (CV F1 = 0.824), which outperformed the other models across all monitored metrics while achieving the lowest variance and the best probability calibration. Conclusions Based on the results obtained, a method for software code defect classification is proposed that combines diversified stacking over feature subspaces with additional decorrelated branches based on Extra Trees and an adaptive logistic regression meta-classifier. The overall quality gain brought the achievable accuracy closer to the theoretical ceiling imposed by label noise in the dataset.