Abstract :

The Ever-growing Size Of Contemporary Software Systems Has Placed Tremendous Demands On Quality Engineering Disciplines, Particularly In The Area Of Automated Fault Localization And Defect Forecasting. The Modern Software Development Environment Is Characterized By Big Code Bases, Swift Release Cycles And Intricate Architectural Dependency Chains, Which Increases The Chances Of Issues Going Undetected In Prerelease. Manual Oversight And Single Static Analysis Technologies Do Not Provide Adequate Coverage For The Large And Intricate Complexity Of Todays Enterprise Systems. The Software Defect Prediction Paradigm Based On Machine-learning Has Gained Popularity As It Helps Software Development Companies To Invest Money In The Statistical Most Probable Problem Areas Of Software Modules In Advance. Although The Improvements Are Measurable, The Single Algorithm Prediction Models Are Fundamentally Flawed In A Number Of Ways: They Do Not Perform Well When The Number Of Training Examples Is Sparse, They Do Not Perform Well When Deployed To A Different Type Of Project, And They Fail To Be Sensitive To The Minority Defect Class When The Class Distribution Is Naturally Skewed. In This Paper, A Novel Two-levels Stacking Structure Of IEM For Software Defect Prediction Problem Is Presented. The First Level Merges Five Different Types Of Base Classifiers Such As DT, RF, Gradient Boosting Machine, Radial Basis Function SVM And Gaussian Naïve Bayes Classifiers Trained Independently On The Pre-processed Software Metrics. The Second Layer Uses A Meta-learner That Is A LR, Which Combines The Probabilistic Output Of The Basic Estimators Through Coefficient Weighting That It Learns, To Get A Polished Final Prediction. The Framework Is A Hybrid Feature Selection Pipeline, Which Combines The PCIRE With The RFE Based On A RF Importance Estimator, To Systematically Shrink The Feature Dimensionality From An Average Of 37 Attributes To 18, While Maintaining Maximum Predictive Information. Class Imbalance Exists In All Defect Datasets And Is Eliminated By Using A Selectively Applied SMOTE Within A Stratiform Ten-fold Crossvalidation Loop, Which Ensures That The Held-out Data Is Not Contaminated By Data In The Training Partitions. The Proposed IEM Is Extensively Empirically Validated On Five NASA PROMISE Repository Benchmarks (CM1, KC1, KC2, PC1, And JM1) And Outperforms All The Competing Baselines Under Testing By Statistically Significant Margins With Mean Classification Accuracy Of 94.3%, F1-Score Of 90.5%, And AUC-ROC Of 0.97.

Published:

10-7-2026

Issue:

Vol. 26 No. 7 (2026)

Page Nos:

409 - 422

Section:

Articles

License:

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

How to Cite