Source selection and transfer defect learning based cross-project defect prediction

Abstract

Software defect is an important metric to evaluate software quality. Too many defects will make the software unavailable and cause economic losses. The aim of software defect prediction (SDP) is to find defects as early as possible. Based on this, source project selection and transfer defect learning based cross-project defect prediction (STCPDP) is proposed. This method firstly sets the threshold of the metrics to predict the defect more effectively, secondly computes the similarity between different project versions to find the appropriate train sets, and finally combines the popular transfer defect learning method TCA + to predict software defects based on the logistic linear regression model. Experimental results show that when the defect probability threshold is about 0.4, STCPDP has better performance based on the F-measure metric, and STCPDP can effectively improve the popular CPDP models.