A novel generative oversampling for software defect prediction.

Area of Science:

• Software Engineering
• Machine Learning
• Data Science

Background:

• Software defect prediction is crucial for efficient testing and cost reduction.
• Imbalanced defect data negatively impacts predictor performance.
• Existing methods struggle with the inherent class imbalance in defect datasets.

Purpose of the Study:

• To propose a novel Generative oversampling-based Software Defect Prediction (GeNSDP) model.
• To address the challenge of imbalanced defect data in software engineering.
• To enhance the accuracy and stability of software defect prediction.

Main Methods:

• Developed GeNSDP, a generative oversampling technique using a lightweight generative model.
• Generated synthetic minority instances to balance the defect dataset.
• Employed a deep network for defect prediction on the oversampled data.

Main Results:

• GeNSDP achieved high performance with an average Area Under the Curve of 99.1% and F-measure of 0.92 on NASA and PROMISE datasets.
• The model demonstrated significant improvement over traditional oversampling methods (ROS, SMOTE, COSTE) by 30.1%.
• GeNSDP outperformed baseline models by 14.1%, with statistical validation (Anova Test) confirming its effectiveness.

Conclusions:

• GeNSDP is an effective approach for handling class imbalance in software defect prediction.
• The proposed model provides stable and accurate defect prediction capabilities.
• This research contributes a robust solution for improving software quality assurance.