An explainable AI framework for enhanced software defect prediction using transformer-assisted boosting

Area of Science:

• Software Engineering
• Machine Learning
• Artificial Intelligence

Background:

• Accurate software defect prediction is crucial for mitigating project delays, cost overruns, and reliability issues.
• Existing machine learning models often lack interpretability, hindering practical application in software quality assurance.

Purpose of the Study:

• To develop and evaluate a novel Transformer Assisted Boosting Framework (TABF) for enhanced software defect prediction.
• To combine the predictive power of XGBoost with the interpretability of Transformer self-attention mechanisms.
• To improve the accuracy and explainability of defect prediction models for practical use.

Main Methods:

• The Transformer Assisted Boosting Framework (TABF) integrates XGBoost with Transformer self-attention.
• Evaluation was performed using the NASA Metrics Data Program (MDP) and Code4Code datasets.
• Key software metrics include cyclomatic complexity, Halstead's properties, and lines of code.
• SHapley Additive exPlanations (SHAP) were employed for feature importance analysis.

Main Results:

• TABF achieved superior performance with AUC scores of 0.95 and ROC of 0.96.
• The framework outperformed classical models like Random Forest (92.5% accuracy) and SVM (94.3% accuracy).
• Lines of code and McCabe's cyclomatic complexity were identified as key defect predictors via SHAP analysis.

Conclusions:

• TABF offers a powerful and interpretable approach to software defect prediction.
• The framework bridges the gap between advanced ML/DL models and practical software quality assurance.
• Insights from TABF aid in defect management, resource allocation, and enhancing overall software reliability.