A Cross-project Defect Prediction Model Using Feature Transfer and Ensemble Learning

Cross-project defect prediction (CPDP) trains the prediction models with existing data from other projects (the source projects) and uses the trained model to predict the target projects. To solve two major problems in CPDP, namely, variability in data distribution and class imbalance, in this paper we raise a CPDP model combining feature transfer and ensemble learning, with two stages of feature transfer and the classification. The feature transfer method is based on Pearson correlation coefficient, which reduces the dimension of feature space and the difference of feature distribution between items. The class imbalance is solved by SMOTE and Voting on both algorithm and data levels. The experimental results on 20 source-target projects show that our method can yield significant improvement on CPDP.