Using Smote for Convalescing Software Defect Prediction

Abstract

Software fault prediction models use software metrics and fault occurrences statistics, collected from previous versions of software products. These models attempt to predict the defect status (probability) of the software components. Developing an accurate and efficient software defect prediction model remains a challenging issue due to existence of outliers in data sets used in defect prediction process and that the unbalanced data sets badly influence the performance of software fault prediction model. Most of the prediction model make use of all the software metrics collected from previous projects , although there is no need to employ all software metrics collected for fault prediction as some of these software metrics are redundant and should not be used to develop prediction model due to curse of dimension. We previously proposed [23] a model to deal with two issues i.e. outlier detection and attribute selection. In this paper, we have improved our previous model and apply Synthetic Minority Over-sampling Technique (SMOTE) on the datasets to deal with class imbalance. We have compared the classification results of both models to validate the strength of improved approach.