Research On Software Defect Prediction Based On Transfer Learning

With the increasing of the scale and complexity in software,the possibility of hiding defects in software is becoming larger and larger.The defects in software may have significant influence on the production and life,so the quality of the software has attracted much attention.During the process of software development,discovering and fixing defects in advance can allocate limited test resources effectively.Software defect prediction(SDP)constructs a defect prediction model by mining source code files and development documents,and predicts new software modules.However,for newly de-veloped software,there is insufficient historical training data.Therefore,cross project defect prediction(CPDP)emerges as the times require.CPDP uses historical data from other project(called source project)to construct an effective defect prediction model to predict whether the software module in current project(called target project)contains defects.There are many distribution differences between the source project data and the target project data,which can seriously affect the performance for CPDP.Therefore,it has significant influence for CPDP to design effective methods to select data which has similar distribution to target project from the source project.In the field of CPDP,in order to reduce the distribution differences between the source project data and the target project data,this article presents a framework,named FeCTra,which is based on feature transfer and instance transfer.Firstly,this frame-work divides all features into clusters,and removes features with large distribution differences,which can successfully perform feature transfer.Then,using TrAdaBoost technology,several base classifiers are trained on the source instances and the target instances,then a powerful integrated prediction model is constructed.By applying FeCTra to the Relink and AEEEM datasets,it can be found that FeCTra can effectively remove features and instances which have large distribution differences and improve the performance of defect prediction.The main contributions of this paper are summarized as follows:1.This paper presents a framework FeCTra for cross project defect prediction based on feature tranfer and instance transfer.In order to find defective file in the process of software development,this paper presents a framework FeCTra for cross project defect prediction which is based on feature transfer and instance transfer.This paper describes in detail the mo-tivation of such method,the knowledge of feature transfer and instance transfer.In addition,the measurement of feature correlation and the specific design of the algorithm are analyzed in detail.2.An empirical study on the effectiveness of FeCTra.In order to verify the effectiveness of the FeCTra framework in cross project defect prediction,this paper makes an empirical study on Relink datasets and AEEEM datasets,comparing the experimental results with other typical meth-ods.The effects of different factors on the performance of FeCTra were studied by different experimental designs.