Ensemble Learning for Software Defect Prediction: Performance, Practicality and Future Directions

Ensemble learning is a leading approach in software defect prediction (SDP), offering improved predictive performance on imbalanced and high-dimensional datasets. Despite growing research interest, persistent gaps remain in model interpretability, generalizability, and reproducibility, limiting its practical adoption. This paper presents a comprehensive analysis of 56 peer-reviewed studies published between 2020 and 2025, spanning both journal and conference venues. Findings show that ensemble methods, especially when combined with sampling, feature selection, or optimisation, consistently outperform single classifiers on important metrics such as F1-score, area under the curve, and Matthew correlation coefficient. Nonetheless, few studies incorporate explainability frameworks, effort-aware evaluation, or cross-project validation. Additionally, most models are static, rely on within-project testing, and depend on legacy datasets such as PROMISE and NASA, which limit external validity. Building on this synthesis, the review highlights future research priorities, including interpretable ensemble architectures, adaptive modelling, dynamic imbalance handling, semantic feature integration, and real-time prediction. Standardised benchmarks, transparent, scalable designs are recommended to bridge the gap between experimental performance and deployment-ready SDP solutions.