Research On Cross Project Software Defect Prediction Based On Feature Transfer And Instance Transfer

With the development of science and technology,the role of software in social activities is increasing,and the reliability of software has received more and more attention from the society.At present,software testing is still the main method to detect software defects.However,in the face of a large-scale software system with complex functions,it is a huge challenge to efficiently complete software defect detection within the specified time.Software defect prediction technology can construct a software defect prediction model to predict new software modules through the historical data information of the software,thereby providing testers with an effective test resource allocation plan and improving the efficiency of software testing.For newly developed software systems,it may not be able to provide sufficient software historical information.Therefore,scholars began to try cross-project software defect prediction research,using historical data information of other software to build defect prediction models to predict the target software.In order to further improve the performance of cross project software defect prediction,this paper proposes a cross project software defect prediction method based on feature transfer and a cross project software defect prediction method based on instance transfer,which reduces the difference of data distribution between projects from the perspective of feature and instance respectively,and improves the accuracy of model prediction.Among them,in feature research,this paper proposes a wrapped feature transfer algorithm WFTG based on genetic algorithm,which uses genetic algorithm to search the optimal transfer feature subset through the verification results of candidate feature subset on some labeled target data.In instance research,this paper proposes a cost sensitive learning based instance transfer algorithm Tr Adacost.By introducing the cost sensitive learning mechanism,the weight of different misclassified cost samples in training is changed,so as to alleviate the impact of class imbalance on model training.Experimental results on AEEEM data set show that the proposed method can achieve better prediction results than the classic cross project software defect prediction method.