An Intelligent Ensemble-Based Framework for Early Software Defect Prediction Using Hybrid Feature Selection

Abstract

Software defect prediction is crucial for enhancing system quality and reducing long-term maintenance costs. Early identification of fault-prone software modules allows development teams to optimize testing efforts and improve post-release stability. Despite the potential of machine learning, existing models often face challenges, including performance instability across datasets, excessive feature noise, and the separation of feature selection from ensemble learning. To address these gaps, this study proposes an intelligent framework that integrates hybrid feature selection with a weighted ensemble mechanism. The hybrid feature selection approach combines correlation-based filtering with Recursive Feature Elimination (RFE) to reduce dimensionality and eliminate irrelevant features. Subsequently, the framework employs a weighted ensemble model based on F1-scores, which dynamically assigns weights to base learners—Random Forest, Support Vector Machine, and Gradient Boosting—to enhance predictive stability. Experimental results across five benchmark datasets (PC1, JM1, KC1, MW1, and CM1) demonstrate that the proposed framework consistently achieves superior predictive performance and stability compared to individual base learners. Statistical validation using the Wilcoxon signed-rank test confirms the significance of the improvements, particularly for larger datasets. By effectively balancing data quality through hybrid feature selection and leveraging diverse learner strengths, this framework provides a robust and practical solution for proactive software quality assurance.