Abstract:

Software fault prediction is an essential tool in improving the quality of software and minimizing the maintenance expenses of the software by detecting modules that were prone to defects at an early stage of the development lifecycle. Nevertheless, the class imbalance, in which faulty modules are a small subset of the dataset, is also a significant issue to the traditional machine learning classifier, which can subsequently result in low detection of instances of minority classes. To eliminate this problem, the current research suggests MC-SMOTE (Meta-Clustered SMOTE with Cleaning), a new hybrid sampling method that combines the clustering-based selective oversampling and mild under sampling with noise removal through ENN and Tomek Links. MC-SMOTE produces quality balanced data, minimizes artificial noise, and stabilizes decision-borders to better detect the minority-classes. The efficiency of the suggested method is measured on 6 NASA PROMISE datasets (CM1, KC1, JM1, PC1, PC2, PC3) with the help of 6 popular classifiers: Random Forest, Naive Bayes, K-Nearest Neighbors, Support Vector Machine, Logistic Regression, and Decision Tree. The experiment has shown that MC-SMOTE demonstrates significantly better results in all metrics, such as Accuracy, Precision, Recall, F1-Score, AUC-ROC, MCC, and G-Mean, and shows significant improvement in minority-class recognition and false alarm reduction. The results indicate that the hybrid methodology provides an increase in the reliability of the prediction of fault and effective generalization of the concept across various classifiers, which provide a strong solution to the challenge of dealing with class imbalance in the context of software quality assurance.

Key words: software fault prediction, class imbalance, hybrid sampling, MC-SMOTE, ENN, Tomek links