Software Defect Prediction using Stacking Ensemble Combined with Oversampling Technique

Despite ongoing advancements in software engineering, software defects pose significant hindrances to improving software quality and reducing the overall cost. Although contemporary studies utilize machine learning for defect prediction, achieving high accuracy with imbalanced data remains a challenge. To enhance the defect prediction, this paper presents a stacking based model combined with oversampling techniques. The proposed model employs a stacking ensemble model comprising of Random Forest, Support Vector Machine, and Extra Gradient Boosting, K-Nearest Neighbors as base classifiers, with Logistic Regression serving as the meta-learner. Additionally, random oversampling is leveraged as an initial step to effectively address data imbalance. Four benchmark datasets from the PROMISE repository (KC1, KC2, PC1, and PC3) are used to evaluate the effectiveness of the proposed approach using accuracy, precision, recall, F-measure and ROC. The results establish the proposed approach significantly improves defect prediction across all datasets and metrics.