Deep Residual Network Voting Ensemble for Software Defect Prediction Enhanced by GAN-Based Data Augmentation
Unknown authors
Abstract
Software Defect Prediction (SDP) is essentially a major factor in improving the quality of software by detecting the modules likely to contain defects during early development during the development lifecycle.Defect datasets are generally found to have imbalances in classes due to the complicated interactions of features, making it ineffective to build any traditional machine learning model on them, as in the case of real-world settings.The current study aimed at proposing a hybrid framework, DRNVE-WGAN-GP, that combines data augmentation with an ensemble of Deep Residual Networks (DRNs) in order to accurately predict defects using Wasserstein loss and gradient penalty (WGAN-GP).The GAN module is applied to synthetically generate samples from the minority class, thus more effectively mitigating class imbalance than traditional sampling methods.A DRN ensemble trained with class weights and optimized through a voting classifier.To ensure statistical robustness, the performance of the model had been measured using the 5-Fold Stratified Cross-Validation method, which aggravates the variance while keeping an equal class distribution into the folds.The study has conducted experiments using the NASA PC1, PC5, KC1, KC2 and JM1 datasets and compared the performance of the model to baseline deep learning models, including Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), towards validating the suggested approach.Synthetic Minority Over-sampling Technique -Tomek Links (SMOTE-Tomek) is also evaluated against the WGAN-GP as a traditional resampling technique for class balancing.The proposed framework achieves an average accuracy of 98.11%, precision of 67.65%, recall of 69%, F1-score of 68.32%, and an Area Under the Curve-Receiver Operating Characteristic (AUC-ROC) of 97.55%, confirming its superiority in predictive performance across all evaluation metrics.