Influence Analysis Method of Class Imbalance on Software Defect Prediction Model Stability and Prediction Performance

The paper proposes a method for analyzing the influence of class imbalance on software defect prediction model stability and prediction performance. Firstly, the original data set is constructed into a set of new data sets whose unbalance rate is less than the original data set's unbalance rate by using the undersampling method. Where, fixed seeds are used in the construction of the data set to ensure that the data in the same unbalanced rate data set constructed by the same data set is the same, so as to reduce the randomness of the results of each run. Secondly, the MCC value is taken as the performance evaluation indicator of the prediction model, and the new data set generated each time is put into the classification algorithm of the model for training and prediction evaluation, so as to obtain the MCC value at different unbalanced rate for the current data set. We also propose a performance stability evaluation indicator. The experimental results show that, MCC is more suitable as the stability evaluation indicator of software defect prediction model under the condition of class imbalance compared with AUC. For the stability of software defect prediction performance, the cost sensitive model performs better than the ensemble model. © 2023 Chinese Institute of Electronics. All rights reserved.