Software Defect Detection and Risk Prediction based Decision Support

Abstract

In order to improve the reliability of software systems by minimizing software testing costs through SDP, it was found that there is a need for it. However, it is observed that current solutions treat each of these issues independently of one another. To resolve this challenge, this paper introduces FA-MRFO-DL that combines CNN for extraction of local features, BiLSTM for learning interactions between metrics, attention-based weighting, XGBoost classifier, FA-MRFO, and SHAP. Evaluated on five NASA MDP benchmark datasets (JM1, KC1, CM1, KC2, PC1; 10,963 instances), the proposed model achieves 97.64% accuracy, 96.65% F1-score, and 0.989 AUC for binary defect detection. For three-class severity prediction, the model achieves 91.36% accuracy (κ = 0.877) on KC1 using native Bugzilla annotations—the most methodologically rigorous evaluation—and 94.29% accuracy (κ = 0.906) under a cross-project label transfer protocol. ADWIN-based drift adaptation sustains 95.21% accuracy under simulated distributional shift (sequential row-order partitioning of JM1 in the absence of native timestamp metadata). All comparisons are statistically confirmed via 5×2cv paired t-test (p < 0.001, Cohen’s d ≥ 2.81). SHAP analysis identifies cyclomatic complexity, coupling, and lines of code as the dominant defect predictors. Results are constrained to procedural C/C++ legacy codebases; generalisation to modern software ecosystems requires further validation.