Training data selection for imbalanced cross-project defect prediction

Abstract

Machine learning methods have been applied in software engineering to effectively predict software defects. Researchers proposed cross-project defect prediction (CPDP) for cases in which few or no data are available. CPDP uses the labeled data of a source project to construct a prediction model for the target project. However, the prediction performance remains inferior because the training data selection for the source project is ineffective. In this paper, the Jensen-Shannon divergence is first applied to automatically select the source project most similar to the target project. Subsequently, a grouped synthetic minority oversampling technique (SMOTE) is applied to improve the class imbalance of the projects. Finally, relative density estimation is performed to select the data for the source project. The experimental results demonstrate that the proposed method improves the prediction performance and exhibits high adaptability to different classifiers.