Adaptive Centre-Weighted Oversampling for Class Imbalance in Software Defect Prediction

In the field of software engineering, software defect prediction can maintain the high quality of software products, which is a popular current research topic. However, class imbalance affects the overall classification accuracy of software defect prediction models which is the key issue to be resolved. A new method called adaptive centre-weighted oversampling (ACWO) is proposed to effectively address imbalanced learning problems. First, an appropriate neighborhood size and neighbors are determined for each minority class sample. Then, for a minority class sample, the adaptive centre that is within its neighborhood range, its neighbors and the minority class sample are used to generate synthetic samples. Finally, oversampling of each minority class sample is carried out based on the weights assigned to them. These weights are obtained according to the neighborhood sizes and Euclidean distances to the centre. Afterwards, the software defect prediction model is eventually established by ACWO algorithm with stacked denoising autoencoder neural network. Experimental results show that the software defect prediction model based on ACWO algorithm has a better performance than based on many existing class imbalance learning algorithms according to the precision P, recall R, F1 measure, G-mean, and AUC values.