A Stacked Regressor for the Prediction of the Number of Defects in Software Modules

Context: Software system releases are often subject to multiple updates leading todifferent versions. These updates are due in part to defects that occur in the development stage. This requires companies to allocate more resources to more defective modules.Objective: This paper proposes a Stacked Regressor (SR) that predicts the numberof defects in a software module based on class-level metrics.Method: We four different base models for the SR: Support Vector Regres-sion (VR), K -Nearest Neighbors (KNN), Random Forest Regressor (RFR) andExtra Trees Regressor (ETR). Two versions of SR are implemented which differ intheir meta-regressor. We test the approach in a within-version and a within-projectsetting.Results: We validate our proposed approach on different software systems that differin size, number of defects and class imbalance and compare it to the base learners.We notice that the SR provides a significant improvement over its base models interms of the precision of the number of defects predicted per module.Conclusion: The proposed approach groups base learners together under a StackedRegressor providing substantial improvements over their individual predictive capabilities.