Explainable Machine Learning for Software Defect Prediction in Enterprise Applications

Abstract

Defect prediction models are now embedded in enterprise delivery pipelines, and post-hoc explainers such as SHAP and LIME are routinely attached to their risk scores. Whether those explanations are faithful to the model, stable across retraining, and useful to the engineers who must act on them is rarely tested. This critical integrative review examines explainable machine learning for software defect prediction across five evidence streams: process versus product metrics, just-in-time (change-level) prediction, class imbalance treatment, effort-aware evaluation, and cross-project transfer. A structured protocol covering digital libraries and standards published through 2025 was used to appraise studies for temporal validity, label quality, explanation evaluation and decision relevance. The synthesis finds that discrimination metrics overstate operational value when time-ordered splits, effort normalisation and label latency are ignored; that feature-attribution rankings shift under resampling, collinearity and retraining; and that developer benefit depends on whether explanations point to modifiable, change-local factors. The review contributes the FACTS-DP framework, which separates explanation fidelity and stability, actionability, cost-aware evaluation, temporal validity and evidence stewardship; six falsifiable propositions with stated tests; and a validation design that pairs rolling-origin evaluation with a developer decision experiment. The propositions are untested. Enterprise teams should report effort-aware yield and explanation stability alongside AUC, and should release explanations to developers only when documented fidelity and stability thresholds are met