Improving Prediction Robustness of VAB-SVM for Cross-Project Defect Prediction

Abstract

Software defect prediction is important for improving software quality. Defect predictors allow software test engineers to focus on defective modules. Cross-Project Defect Prediction (CPDP) uses data from other companies to build defect predictors. However, outliers may lower prediction accuracy. In this study, we propose a transfer learning based model called VAB-SVM for CPDP robust in handling outliers. Notably, this method deals with the class imbalance problem which may decrease the prediction accuracy. Our proposed method computes similarity weights of the training data based on the test data. Such weights are applied to Boosting algorithm considering the class imbalance. VAB-SVM outperformed the previous research more than 10% and showed a sufficient robustness regardless of the ratio of outliers.