Research On Software Defect Prediction Method Based On Metric Compensation And Semantic Feature Optimization

In the current era of evolving information technology,as the scale and complexity of software continue to increase,the pursuit of quality and efficiency by technical staff in the software development process is also increasing,so there is much concern about how to ensure the quality of software.Software defect prediction,as a preventive technique that identifies potential defects early in the software development cycle,helps to identify and deal with potential defects as early as possible in software development requirements analysis,design,coding and testing,thus achieving efficient development and high quality software.Traditional machine learning-based prediction techniques typically build defect prediction models using manually designed software metrics,but traditional manual features often fail to capture structural semantics and data flow information in the source code.Deep learning techniques are used to automatically learn the semantic representation of a program in order to bridge the gap between the semantic information of the program and the defect features when performing defect prediction.However,existing defect prediction methods usually only consider traditional features or a single type of semantic features,or fail to learn effective feature representations in the process of obtaining semantic features from software source code,which has an impact on the performance of the model.Based on this,this study presents two different approaches to building prediction models from the perspectives of traditional features and joint features,and proposes two different approaches to software defect prediction.The main contents and contributions of the study are as follows:1.In response to the problems of feature redundancy and data discrepancies in traditional cross-project software defect prediction methods,a two-way metric compensation method(pe Up Me Com)based on transfer learning and feature selection is proposed.The method uses Pearson correlation coefficients to numerically analyse the correlation between various metrics and defect categories,filters the irrelevant and redundant features,and then uses Transfer Component Analysis(TCA)to find a mapping space for the data features of the source and target projects,and finally makes a Two-way adaptation to further enhance the similarity of the data distribution.The experimental comparison with various Cross-Project Defect Prediction(CPDP)methods shows that the defect prediction model constructed using pe Up Me Com method has higher AUC and F1-measure values on experimental datasets such as NASA.2.To address the problem that existing CPDP methods have difficulty in obtaining valid feature representations in the program code when building defect prediction models,a software defect prediction method based on semantic feature enhancement(SFE-DP)is proposed.To find a balance between traditional manual features and semantic features to allow the model to learn the weighting ratio of the two types of features adaptively,the method performs data augmentation before reconstructing the output of semantic information in the process of extracting semantic features.The model uses a self-attentive mechanism as well as a matching layer to capture key features in the program semantics and fuse traditional features in an appropriate ratio.The SFE-DP method is compared with various deep learning based CPDP methods and is found to give better results in terms of model classification performance.3.A prototype system of software defect prediction based on metric compensation and semantic feature optimization is designed and implemented.The system mainly includes defect prediction method pre-processing module,defect prediction method execution module,and prediction result statistical analysis module.The defect prediction method execution module mainly integrates the two proposed defect prediction methods and related comparison methods,and the system automates the prediction process of each algorithm and confirms the effectiveness of the methods.