Software Defect Prediction: A Systematic Review and a Conceptual Framework for Stable, Actionable Explanations (AISD-F)
Abstract
Machine learning-based Software Defect Prediction (SDP) has largely focused on improving prediction accuracy, but recent research has placed greater emphasis on making predictions understandable and useful to developers. This systematic review examines the use of Explainable Artificial Intelligence (XAI) in SDP and the evidence supporting its reliability in practice. Following PRISMA guidelines, 40 peer-reviewed studies, primarily published between 2020 and 2025, were selected from IEEE Xplore, SpringerLink, ScienceDirect, Google Scholar, and Scopus. The studies were examined in terms of XAI methods, explanation reliability and stability, dataset quality, and model selection. LIME and SHAP were the most frequently used methods, yet multi-run evaluations showed changes in their feature rankings across configurations, in contrast to the reliability reported by many single-run studies. Data-quality problems, particularly SZZ-related label noise, class imbalance, and multicollinearity, often had a greater effect on prediction reliability than model architecture. The review also found growing use of line-level and just-in-time prediction, although consistent validation at industrial scale remains limited. Within the reviewed corpus, no study provides controlled evidence that XAI explanations improve developers' ability to resolve software defects, and the field currently lacks industrial-scale validation and controlled developer studies of XAI-assisted defect resolution. Based on these findings, this paper proposes the Actionable and Interpretable Software Defect Framework (AISD-F), a conceptual, five-stage modular architecture (not yet empirically validated) that is intended to integrate data quality assurance, hybrid model selection, dual-layer explanation, and CI/CD into a unified research agenda for explainable SDP, rather than a deployment-ready solution.
Related
- Towards a conceptual clarification and possible applications of Meehl’s crud factor: A preregistered systematic review
- Explanations, Prompts, and Formalizations: Arguments for New Norms in LLM-Enabled Mathematical Research
- Observable Patterns Are Not Explanations: A Causal-Geometric Analysis of Latent Reasoning Models
- When Are Concept Bottleneck Model Explanations Faithful and Compact?
- Decision-Tree Surrogates for Weak Counterfactual Explanations in Integer Optimization
BibTeX
@article{thubisi2026software,
title = {{Software Defect Prediction: A Systematic Review and a Conceptual Framework for Stable, Actionable Explanations (AISD-F)}},
author = {Dikeledi Irene Thubisi and Bassey Isong and Rodney Sebopelo},
journal = {Journal of Information Systems and Informatics},
year = {2026},
month = oct,
doi = {10.63158/journalisi.v8i5.1823},
url = {https://doi.org/10.63158/journalisi.v8i5.1823},
note = {CC BY},
}
RIS
TY - JOUR
TI - Software Defect Prediction: A Systematic Review and a Conceptual Framework for Stable, Actionable Explanations (AISD-F)
AU - Dikeledi Irene Thubisi
AU - Bassey Isong
AU - Rodney Sebopelo
PY - 2026
DA - 2026/10/24
JO - Journal of Information Systems and Informatics
DO - 10.63158/journalisi.v8i5.1823
UR - https://doi.org/10.63158/journalisi.v8i5.1823
AB - Machine learning-based Software Defect Prediction (SDP) has largely focused on improving prediction accuracy, but recent research has placed greater emphasis on making predictions understandable and useful to developers. This systematic review examines the use of Explainable Artificial Intelligence (XAI) in SDP and the evidence supporting its reliability in practice. Following PRISMA guidelines, 40 peer-reviewed studies, primarily published between 2020 and 2025, were selected from IEEE Xplore, SpringerLink, ScienceDirect, Google Scholar, and Scopus. The studies were examined in terms of XAI methods, explanation reliability and stability, dataset quality, and model selection. LIME and SHAP were the most frequently used methods, yet multi-run evaluations showed changes in their feature rankings across configurations, in contrast to the reliability reported by many single-run studies. Data-quality problems, particularly SZZ-related label noise, class imbalance, and multicollinearity, often had a greater effect on prediction reliability than model architecture. The review also found growing use of line-level and just-in-time prediction, although consistent validation at industrial scale remains limited. Within the reviewed corpus, no study provides controlled evidence that XAI explanations improve developers' ability to resolve software defects, and the field currently lacks industrial-scale validation and controlled developer studies of XAI-assisted defect resolution. Based on these findings, this paper proposes the Actionable and Interpretable Software Defect Framework (AISD-F), a conceptual, five-stage modular architecture (not yet empirically validated) that is intended to integrate data quality assurance, hybrid model selection, dual-layer explanation, and CI/CD into a unified research agenda for explainable SDP, rather than a deployment-ready solution.
SN - 2656-5935
SN - 2656-4882
ER -