Integrated Machine-Learning Pipeline for Software Defect Prediction in Resource-Constrained Environments

Software defect prediction (SDP) enhances software reliability under settings with very scarce testing resources, although most literature tends to model preprocessing, class-imbalance surpass, feature selection and modelling as independent phases, thus impeding any realistic deployment. This paper presents a single, end-to-end machine-learning pipeline, which combines all these steps into one framework. Four baselines and a stacked hybrid ensemble were studied and measured under equal experimental conditions which were controlled. Experimental findings have proved that ensemble and hybrid pipeline developably outperform baseline performance on every imbalance-sensitive measure in within-project and crossproject settings. Compared to cross-project performance, the ensembles have high robustness, thus justifying the use of integrated ensemble pipelines as the practical way to address software development environments within limited testing resources.