Research On Software Defect Prediction Based On Improved Balanced Distribution Adaption Algorithm

The application of portable devices meets people's various needs.The software that supports such needs has also developed rapidly.Then the scale and complexity of existing software is constantly increasing,which also makes more bugs in software possible.Software defect prediction uses historical data to train a model to achieve the purpose of defect prediction.When a project does not have historically marked defect data,cross-project defect prediction is an effective way to solve this question and improve software product quality.Currently,crossproject defect prediction methods based on machine learning have attracted attention to researchers.This article will improve the performance of defect prediction system by introducing transfer learning technology.However,the existing methods often have these two problems:(1)they mainly focus on the marginal distribution differences in the process of transferring data,and ignore the conditional distribution differences,which will lead to unsatisfactory prediction performance;(2)the data set may be highly imbalance,while using imbalanced data can affect defect prediction performance.For the above two questions,this paper use a two-step transfer by kernel method and transfer learning.First,the kernel method is used to map low-dimensional features to high-dimensional features space.Then,balanced distribution adaptation in transfer learning is used to balance the marginal distribution and conditional distribution.At the same time,the class imbalance data will resample and assign different weights to the samples based on the Euclidean distance.This paper proposes a BDA+ model based on class imbalance processing and balance distribution adaptation.In order to evaluate the effectiveness of BDA+in cross-project defect prediction,this paper uses four performance measures(e.g.F-measure,g-mean,Balance,and AUC)to experimentally verify 18 projects from four public data sets.The experimental results show that compared with 10 basic methods,BDA+,achieves better performance.Compared with the 6 traditional methods,the F-measure,g-mean,AUC,and Balance are improved: 22.78%-39.29%,7.61%-19.66%,12.36%-26.19%,5.60%-18.75%?...