A Hybrid Feature Selection and Ensemble Learning Framework for Software Defect Prediction with Class Imbalance Handling

Abstract

Software defect prediction (SDP) is essential for improving software quality since it finds error-prone modules early in the development lifecycle. Current methods produce inflated and erroneous performance metrics because of data leaks, inadequate class imbalance management, and reliance on antiquated classifiers. By integrating a soft voting ensemble of six contemporary gradient boosting classifiers XGBoost, LightGBM, CatBoost, Random Forest, Extra Trees, and Gradient Boosting with Mutual Information filtering, this study provides a MI-based feature selection technique that gets beyond these restrictions. To avoid data leakage, SMOTE is only used on training data to address class imbalance. Five publicly accessible PROMISE repository datasets CM1, PC1, KC1, KC2, and JM1 are used in the experiments. For all five datasets, the average accuracy is 82.86%, while the average F1-score and ROC-AUC are 46.92% and 0.778, respectively. Previous studies on the same datasets only give accuracy numbers of up to 94% for imbalanced defect datasets, when the number of non-faulty modules significantly exceeds that of defective ones. This is a deceptive figure that hides insufficient ability to detect defects. F1-score, Precision, Recall, and ROC-AUC are used to measure performance; these metrics offer a more accurate evaluation than accuracy alone. The results show that MI-based feature selection improves prediction performance while lowering computational overhead; on all five datasets, contemporary ensemble techniques regularly beat traditional classifiers.