An Advanced Stacked Ensemble Framework for Enhanced Software Bug Predictions

Abstract

Software Bug Prediction (SBP) is a critical activity in software engineering that seeks to identify defect-prone modules before release, enabling focused testing and reducing post-release failures. With the greater scale and complexity of software, data-driven prediction techniques are necessary to maintain reliability and limit maintenance costs. Machine learning methods have led to recent advances in SBP by inducing defect patterns from historical data. However, existing models struggle to address class imbalance, feature redundancy, and limited generalization across projects. Furthermore, most previous studies focus on single classifiers, which are unable to capture complex data patterns effectively. These limitations reduce the utility of the real world and make it difficult to maintain consistent cross-project performance. To address these issues, we propose a new stacked ensemble framework that combines three base learners (Random Forest, XGBoost, CatBoost) and a Multi-Layer Perceptron model as a meta-learner. We performed a systematic ablation study on four preprocessing pipelines (SMOTE+RFE, SMOTE+Boruta, ADASYN+RFE, ADASYN+Boruta) to evaluate the state-of-the-art model. Empirical evaluation across ten benchmark datasets from the NASA and PROMISE repositories shows consistent gains over individual classifiers and competitive literature baselines, with the propose framework achieving a maximum accuracy of 96.52% while providing notable improvements in bug-prediction robustness.