A Unified Ensemble–Deep Learning Framework for Software Defect Prediction: Boosting/Voting Meets CNN–RNN Modeling

Software defect prediction remains a cornerstone of modern software reliability engineering, enabling teams to proactively allocate testing effort, prioritize code reviews, and mitigate operational risk. Despite decades of progress, two persistent limitations motivate renewed attention: (i) conventional metric based predictors often struggle under dataset shift, noisy labels, and evolving development practices, and (ii) deep learning approaches that ingest code tokens or learned representations may be data hungry, difficult to calibrate, and challenging to operationalize in enterprise pipelines that require transparency and governance. This manuscript presents a unified framework that explicitly merges (a) ensemble machine learning—via boosting and voting across heterogeneous metric learners—with (b) a CNN–RNN representation learner designed to capture local and sequential defect cues. The proposed framework introduces a reliability aware fusion layer that calibrates probabilities, quantifies uncertainty, and performs cost sensitive thresholding to align predictions with quality of service objectives. Beyond modeling, the manuscript provides an end to end blueprint for integrating defect prediction into cloud native delivery workflows, including feature lineage, monitoring hooks, and explainability artifacts suitable for high stakes environments. A worked example illustrates how boosted metric learners and CNN–RNN predictors can be combined through weighted soft voting and stacking to produce stable risk scores. The framework is designed to support both within project and cross project evaluation regimes and to remain robust when codebases undergo modernization (e.g., monolith to microservices migrations) or platform transitions (e.g., OpenShift adoption).