Class Imbalance Data-Generation for Software Defect Prediction

Abstract

The imbalanced nature of class in software defect data, which including intra-class imbalance and inter-classes imbalance, increases the difficulty of learning an effective defect prediction model. Most of sampling and example generation approaches just focused on inter-class imbalanced defect data, and they are not effective to handle the issue of intra-class imbalance. This paper proposed a distribution based data generation approach for software defect prediction to deal with inter-class and intra-class imbalanced data simultaneously. First, the classified sub-regions are clustered according to the distribution in the sample feature space. Second, the data are generated by corresponding strategies according to different distribution in sub-regions, where the inter-class balance is achieved by increasing the number of defective samples, and the intra-class balance is achieved by generating different density of data in different sub-regions. Experiment results show that the proposed method can reduce the impact of data imbalance on defect prediction and improve the accuracy of software defect prediction model effectively by generating inter-class and intra-class balanced defects data.