KCO: Balancing class distribution in just-in-time software defect prediction using kernel crossover oversampling

Area of Science:

• Software Engineering
• Machine Learning
• Data Science

Background:

• Defect prediction models are crucial for software quality.
• Imbalanced datasets pose challenges for model performance.
• Existing resampling methods often fail to address data redundancy and noise.

Purpose of the Study:

• To introduce Kernel Crossover Oversampling (KCO), a novel oversampling technique.
• To enhance defect prediction by generating diverse and balanced datasets.
• To mitigate redundancy and noise in imbalanced datasets.

Main Methods:

• Kernel Principal Component Analysis (KPCA) for dimensionality reduction.
• Spectral clustering for identifying optimal interpolation regions.
• Crossover interpolation to generate synthetic defect data.

Main Results:

• KCO consistently achieved F-scores between 21% and 63% across six datasets.
• KCO demonstrated superior prediction performance compared to baseline techniques.
• Significant improvements were observed in both within-project and cross-project defect predictions.

Conclusions:

• KCO effectively addresses data imbalance, redundancy, and noise in defect prediction.
• The proposed technique enhances the accuracy and reliability of defect prediction models.
• KCO offers a promising approach for improving software development quality.