| Περίληψη: |
Detecting software defects in large-scale systems is critical to minimizing failures and ensuring software reliability, as emphasized by the Standish Group's Chaos Report. In sectors like finance, healthcare, aerospace, and enterprise software, unnoticed defects can result in security vulnerabilities, system malfunctions, and substantial economic setbacks, highlighting the necessity for effective defect prediction models. The challenge is compounded by imbalanced datasets and the need for effective feature selection techniques. This study introduces a six-stage frame-work designed to address these challenges: data preprocessing, handling class imbalance, feature selection, ensemble classifier implementation, hyperparameter tuning, and performance evaluation. The framework integrates Mutual Information, Recursive Feature Elimination, and Principal Component Analysis (PCA), incorporating Polynomial Features and scaling to reduce noise. The methodology employs Intelligent Heterogeneous Ensemble Learning (IHEL), combining "Stacking and Voting" with base classifiers like SVM, Logistic Regression, XGBoost, CatBoost, and Gradient Boosting, and optimizing the meta-classifier, Random Forest, through GridSearchCV. Tests conducted on PROMISE datasets (Pc1, Cm1) demonstrate significant improvements in metrics crucial for imbalanced classes, such as Positive Predicted Value, True Positive Rate, Precision-Recall Harmonic Mean, and Sensitivity. The results highlight that the proposed framework outperforms individual classifiers and traditional ensemble models, providing a robust and efficient solution for defect prediction in imbalanced scenarios. This study concludes that adopting this integrated approach can effectively enhance defect detection capabilities in software engineering. [ABSTRACT FROM AUTHOR] |