Unbalanced data processing for software defect prediction

How to solve the imbalance of defect classification in software defect prediction and improve the accuracy of prediction is an important problem in software testing. Thus many machining learning based model, such as self-adaptive Robust Synthetic Minority Over-Sampling Technique (RSMOTE), ware presented for software defect prediction. However, the imbalanced data distribution limited the prediction performance. Addressing to this issue, a RSMOTE-based Data Imbalance Processing (RDIP) model is presented in this paper. Specifically, the normalized outlier data is removed according to the European distance between points in data denoising, and then the fuzzy membership and fuzzy labels of each point are calculated using the Computational Class Fuzzy Algorithm (FCMD), which removes the hazard points and noise points according to the selection boundary point algorithm (BRS). Experimental results the date sets of NASA, Promise show that the average F1-measure of software defect prediction method for data imbalance is 6.98% higher than other comparison algorithms, which can effectively solve the problem of defect classification imbalance in software defect prediction.