Cross-Project Defect Prediction Based On Transfer Learning

With the rapid development of computer technology and the continuous increase in the scale and complexity of software,the possibility of introducing defects in software is increasing.The software defects may cause the software to crush,and even make people's lives and properties in danger.Before the software is released,discovering defects in advance can allocate limited resources and development costs,and improve software quality.Software defect prediction constructs a defect prediction model by mining historical defect data set.However,for newly developed software,there is usually insufficient defect data to build defect prediction model.Therefore,cross-project defect prediction has been developed.It uses historical defect data of source project to predict the defect orientation of target project.In recent years,cross-project defect prediction based on transfer learning has attracted attention of researches in software engineering.The introduction of the transfer learning has greatly reduced the differences between source and target project.However,there are still some problems in the existing cross-project defect prediction methods: 1)At present,most cross-project defect prediction studies are conducted in homogeneous space.For some newly-developed projects,it may be difficult to find defect data with the same software metrics so that defect prediction model cannot be constructed.2)Most defect data are unbalanced.Using unbalanced data to predict defects will decrease the performance of prediction model.3)Most existing defect prediction methods use single source data set.It will be seriously affected the prediction results when the correlation between source and target project is weak.By analyzing and researching the above problems,we present novel defect prediction models for cross-project defect prediction.The contents are described as below:(1)We propose a novel defect prediction model based on instance transfer learning.Based on the TNB approach,we present two improved methods.One is the unified metricrepresentation TNB,UMR_TNB method.The method represents the heterogeneous datasets into the same software metric space,it effectively eliminates the differences between heterogeneous datasets.The experiment results show that the UMR_TNB is suitable for heterogeneous cross-project defect prediction.The other improved method is multi-source TNB,Multi_TNB.To increase the number of source instances which have high correlation with target project,we use k-means to merge and filter multiple source projects.The experiment results show that the performance of defect prediction model which is constructed by using multi-source projects is significantly better than that of using only one source project.It verifies the multi-source adaptability of method.(2)We propose a novel defect prediction model based on feature transfer learning.Based on transfer learning,we propose the HDMP method which is apply to heterogeneous defect prediction.First,using class imbalance method to preprocess source project.Second,using transfer learning method MOMAP to process the heterogeneous source and target project.Transforming the source project into the target space,which make maximize the correlation between the converted source and target project.Finally,using the classification method to predict target project.The experiment results show that the performance of defect prediction of HDMP is obviously better than the existing cross-project defect prediction methods,and the performance is better when using multiple source projects.In summary,this paper aims to use transfer learning to solve practical problems in software defect prediction,which can enrich the application of transfer learning,provide new guide for software defect prediction and is of great significance in improving software reliability.