Enhancing Software Defect Prediction: A Stacking Ensemble Framework with Advanced Feature Engineering

Software defect prediction is a critical aspect of software quality assurance, aiming to identify faulty modules early in the development lifecycle to mitigate potential risks and reduce maintenance costs. This paper presents a comprehensive stacking ensemble framework that amalgamates multiple machine learning techniques to enhance defect prediction accuracy. The methodology encompasses meticulous data preprocessing, a two-stage feature selection process involving Minimum Redundancy Maximum Relevance (mRMR) and polynomial feature expansion, followed by dimensionality reduction using Principal Component Analysis (PCA). Four diverse base learners - Extreme Learning Machine (ELM), Support Vector Machine (SVM), Random Forest, and XGBoost - are trained on the transformed feature set, and their outputs are integrated through a Logistic Regression meta-learner. Empirical evaluations conducted on five benchmark NASA datasets - MC1, CM1, KC2, KC3, and PC1 - demonstrate the robustness of the proposed framework, achieving accuracies up to 93.80%. These results underscore the efficacy of leveraging ensemble learning and sophisticated feature engineering in capturing intricate data patterns for improved software defect detection.