Evaluating Software Quality Through Automated Testing and Predictive Defect Detection Using Machine Learning

Authors

  • Finale Doshi-Velez John A. Paulson School of Engineering and Applied Sciences, Harvard University, Cambridge, USA

DOI:

https://doi.org/10.71465/csb228

Keywords:

Software Quality Assurance, Automated Testing, Defect Prediction, Machine Learning, Continuous Integration

Abstract

Software quality assurance increasingly depends on the ability to combine automated testing with data-driven estimates of defect risk. Conventional test suites reveal failures in executed scenarios but do not necessarily identify which untested changes are most likely to introduce faults. Machine learning models can prioritize code review and testing by learning associations between historical defect outcomes, code structure, change activity, and process characteristics. This narrative review synthesizes established research on software defect prediction and proposes a transparent evaluation framework integrating data preparation, temporal validation, calibrated prediction, risk-based test selection, and continuous monitoring. It distinguishes test effectiveness from model discrimination and emphasizes precision–recall analysis, operational cost, data leakage prevention, and human oversight. A conceptual graph illustrates how testing effort might be redistributed according to predicted risk; it does not represent measured experimental outcomes. The paper argues that machine learning should supplement rather than replace automated tests, and that deployment claims require prospective, project-specific evidence. Limitations include incomplete defect labels, shifting development practices, class imbalance, and uncertain cross-project transferability

References

Software quality assurance increasingly depends on the ability to combine automated testing with data-driven estimates of defect risk. Conventional test suites reveal failures in executed scenarios but do not necessarily identify which untested changes are most likely to introduce faults. Machine learning models can prioritize code review and testing by learning associations between historical defect outcomes, code structure, change activity, and process characteristics. This narrative review synthesizes established research on software defect prediction and proposes a transparent evaluation framework integrating data preparation, temporal validation, calibrated prediction, risk-based test selection, and continuous monitoring. It distinguishes test effectiveness from model discrimination and emphasizes precision–recall analysis, operational cost, data leakage prevention, and human oversight. A conceptual graph illustrates how testing effort might be redistributed according to predicted risk; it does not represent measured experimental outcomes. The paper argues that machine learning should supplement rather than replace automated tests, and that deployment claims require prospective, project-specific evidence. Limitations include incomplete defect labels, shifting development practices, class imbalance, and uncertain cross-project transferability

Downloads

Published

2026-12-31

Issue

Section

Research Articles