Improving Cross-Project Defect Prediction with Weighted Software Modules via Transfer Learning

Abstract

Cross-project defect prediction is investigated to resolve the trouble that software program defect prediction besides historic data. However, there are differences in distribution of software metrics of different software projects, which decreases the overall performance of cross-project defect prediction. This research presents a transfer learning-based cross-project fault prediction approach. The weights of the training software modules are determined by analogy with gravity and compared to those in the test set. The costs associated with different prediction errors are considered. Cost-sensitive C4.5 is employed on weighted training data for cross-project defect prediction. 10 projects from NASA are used in our experiments. Our suggested technique delivers promising cross-project defect prediction results, according to the results, 0.81, 0.41 and 0.8 in average PD value, F-measure value and AUC value, which are best compared to Naïve Bayes based cross-project defect prediction.