A novel generative oversampling for software defect prediction

Software defect prediction is one of the highly active research areas as it allows to focus the testing efforts on the defective modules and reduce the cost of development. The imbalanced nature of defect data poses a threat to the performance of software defect predictors. This study proposes a novel Generative oversampling-based Software Defect Prediction naming GeNSDP. It oversamples defect data by generating synthetic minority instances utilizing lightweight generative model, then the oversampled data is used for defect prediction using deep network. NASA and PROMISE datasets are used for experimentation, for which GeNSDP achieves a remarkable average score of 99.1% for Area Under the Curve, and 0.92 for F-measure. The proposed model outperforms the traditional oversampling methods (including ROS, SMOTE, COSTE) by 30.1%, and selected baseline models by 14.1%. From the statistical evidence obtained by conducting the Anova Test with Bonferroni Post-hoc test at the confidence level of 95%, it is concluded that the proposed model is effective to handle class imbalance and achieve stable defect prediction.