Defect Prediction of Cross Projects Using PCA and Ensemble Learning Approach

Abstract

Cross-project defect prediction (CPDP) is a technique of detecting defects in software modules in which the training and the testing projects for the classification model are different. The effective prediction leads to a more reliable software. The merging of dataset from varying sources results to an imbalanced dataset. The complex structure and the imbalance data make it a challenge for an effective cross-project defect prediction. To overcome these issues, in this paper, we propose a cross-project defect prediction framework. In the first stage of this framework, PCA is applied for dimensionality reduction of the dataset into two components. In the second phase, SMOTE technique of data sampling is applied to handle the class imbalance problem. Then the ensemble classifiers random forest and XGBoost are applied for an effective defect-prediction model. We have conducted the experiments on eight open source software projects. The results are compared with few baseline techniques. The results indicate that the proposed framework gave comparable performance of cross-project defect prediction to some baseline methods.