Cross-project Software Defect Prediction Based On Feature Transfer

With the continuous generation of a large number of software systems,software defect issues have increasingly become the focus of researchers.Software defects are important factors affecting software quality,which may arise from various stages of software development and maintenance.If software products are defective,they may cause huge losses in the actual operation of software products.If you can find defects in the software early,you can reduce the loss as much as possible.Software defect prediction is mainly based on historical data to predict potential defects in the software.It helps to improve software quality.The early software defect prediction mainly uses the existing defect data of the project to establish a prediction model for defect prediction of subsequent software.However,for some emerging projects,historical data is very scarce,even without historical data,then it is obviously impossible to predict the defects within the project.Cross-project software defect prediction,using software defect data from different projects for defect prediction of current projects.This method is a good solution to the problem that the defect prediction data in the project is insufficient and the defect prediction cannot be performed.At the same time,it also introduces new problems for the defect prediction process.Different projects may have distribution differences,classification errors caused by irrelevant information or data redundancy,and the impact of class imbalance problems on prediction results.This paper starts from the problem of data set distribution difference between different projects in the cross-project defect prediction process,improves the various stages of cross-project defect prediction,and combines the migration learning method to propose cross-project software defect prediction based on heterogeneous metrics and based on isomorphic metrics.A cross-project software defect prediction solution designed to further improve cross-project defect prediction performance.The specific research contents are as follows:(1)Cross-project software defect prediction based on heterogeneous metrics: For the cross-project software defect prediction process of heterogeneous metrics,firstly,using multiple feature screening methods to select features for source and target projects respectively,and obtain optimal features.Then,randomly select the appropriate amount of target project data and source project data to mix and sample processing,suppress the class imbalance problem in the source project(training data);use the data enhanced adaptive classification algorithm(Ada MEC algorithm)to train the model,reduce The difference in distribution between the source project and the target project.Finally,the verification experiment of method validity is carried out.The results show that the method can effectively solve the problems arising from the cross-project defect prediction process.(2)Cross-project software defect prediction based on isomorphic metrics: For the cross-project software defect prediction process of homogeneous metrics,firstly,the source project data is sampled to suppress the class imbalance problem in the source project(training data);Using a modified multi-layer adaptive neural network for model training,the MK-MMD metric is added to the target project training loss function,so that the neural network continuously reduces the distribution difference between the source project and the target project.In order to verify the effectiveness of the method,this method is compared with the classical cross-project defect prediction method.The experimental results show that the performance of the proposed method is better than other methods.