Improving Classifier Performance on Imbalanced Software Defect Datasets Using Resampling Techniques

The accurate prediction of software defects is essential for enhancing detecting high-risk areas early in the development cycle to improve software quality. However, many software defect datasets suffer from class imbalance, where the performance of ML models suffers because the number of non-defective examples greatly exceeds the number of defective ones. This study explores the impact of various resampling methods SMOTE, ADASYN, and ROS in addressing this imbalance and improving defect prediction in two imbalanced datasets, PC2 and PC3. We assess the performance of five common machine learning classifiers, such as Random Forest, Decision Tree, and KNN Support Vector Classifier (SVC), and AdaBoost Classifier, under both normal and resampled conditions. Our findings demonstrate that resampling techniques notably improve the recall and F1-score for the defective class, resulting in more robust and accurate predictions. Among the resampling techniques, SMOTE and ADASYN consistently yielded better performance than ROS. The findings highlight the importance of addressing class disparity in the forecast of software defects and demonstrate the potential of resampling approaches in improving model performance, especially for minority classes. This works reviews and compares these techniques in detail and their use in practice in software engineering.