Research On Software Defect Prediction Model Based On Transfer Learning

Software defect prediction technology can be used to predict the existence of software defects and the number of possible defects to determine whether software can be delivered.It is of great significance for software performance improvement,quality assurance and control cost control.Most of the existing software defect prediction models use machine learning algorithms for training and construction.The traditional machine learning applications have inherent limitations such as lack of original data and independent and identical distribution between data.These constraints restrict the development of this technology..transfer learning can better compensate for this shortcoming.Therefore,this topic has carried out in-depth research and optimization on transfer learning and defect prediction related algorithms.Firstly,this paper studies the basic theories and related technologies of transfer learning classification and software defect prediction,and focuses on the analysis and research of existing software defect distribution prediction techniques for software defect prediction problems across projects.The key elements of the optimization process are analyzed for the core steps of the related technology,including the feature selection in the data preprocessing stage and the related algorithms for constructing the prediction model.Then in the process of software prediction model construction,the data preprocessing technology is researched and analyzed firstly,and the research work is carried out based on the filtering feature selection method and SMOTE sampling method.Then,for the two algorithms TrAdaboost and Rareboost,the paper analyzes the sample weight change and sub-classifier construction method,and designs a new TrRareboost algorithm for the unbalanced characteristics of software defect problems,and carries out its rationality and correctness.Analysis.Combined with the Adaboost algorithm,the weight change parameters of TrAdaboost are optimized and a new transfer learning method is obtained.Finally,by designing a series of experiments and testing multiple data sets to verify the superiority of the optimized TrRareboost algorithm compared to TrAdaboost,the existing transfer learning algorithm also includes feature selection and data sampling.The impact of the prediction results,and the comparison of TrRareboost and NN-filter transfer learning methods.The results of these experiments also verify that the TrRareboost algorithm can effectively improve the performance of the software defect prediction model.This paper provides a new algorithm and idea for cross-project software defect prediction,which effectively improves the usability of software defect prediction model,and has good application value,which contributes to the application of this technology to practical problems.