Random Forest and Novel Under-Sampling Strategy for Data Imbalance in Software Defect Prediction

Abstract

Data imbalance is one among characteristics of software quality data sets that can have a negative effect on the performance of software defect prediction models. This study proposed an alternative to random under-sampling strategy by using only a subset of non-defective data which have been calculated as having biggest distance value to the centroid of defective data. Combined with random forest classification, the proposed method outperformed both the random under-sampling and non-sampling method on the basis of accuracy, AUC, f-measure, and true positive rate performance measures.