Cross-Project Logistics Software Defect Prediction Based On Adversarial Discriminative Memory Neural Network

With the rapid development of information technology and the growing expansion of logistics market,logistics software plays an increasingly important role in logistics information management.However,the existence of potential software defects poses a serious threat to the safety and reliability of logistics software.In order to ensure the quality of logistics software and avoid the harm of software defects,software defect prediction has become a popular research direction in the field of logistics software engineering,that is,to build a predictive model of machine learning based on software history data to identify the location of software defects in advance.Since actual development projects often lack sufficient historical data,researchers have proposed cross-project defect prediction by borrowing data from other mature software projects to build prediction models.There are still some problems in the research of cross-project defect prediction,and this paper addresses two problems that have a significant impact on the prediction performance,namely,the lack of semantic and contextual features and the difference in data distribution.The main research of this paper includes the following three parts.(1)To address the problem of missing semantic and contextual features,this paper proposes an abstract continuous long-short term memory neural network(AC-LSTM)software defect prediction model based on deep learning,which aims to extract semantic and contextual features directly from the code level using deep learning techniques.The model uses the abstract syntax tree representation and word vector encoding in the preprocessing stage to convert the code into numerical vectors while maximizing the preservation of semantic and contextual information;and uses the optimized BLSTM more suitable for code sequences in the feature extraction stage to extract the information-rich semantic and contextual features.The model is compared in cross-version experiments on 10 source-target project data pairs related to logistics software development in the PROMISE open source dataset,and the experimental results demonstrate the effectiveness of the AC-LSTM model.(2)To address the problem of data distribution differences,this paper proposes an adversarial learning-based adversarial discriminative memory neural network(GLSTM)cross-project defect prediction model based on the AC-LSTM model,which aims to eliminate the data distribution differences between the source and target domains by adopting the adversarial game idea in adversarial learning.The model applies the structure of GAN in adversarial learning,where the feature vectors of the source and target projects are input into the discriminator as true and false samples for gaming,and the features of the target project are continuously transferred by adversarial training of the target feature extractor and discriminator until the feature distribution of the target project is similar to that of the source project.At this point,the trained classification model of the source project can be directly used to predict the defect propensity of the target project’s feature vectors.The model is used to conduct crossproject comparison experiments on 15 source-target project data pairs related to logistics software development in the PROMISE open source dataset,and the experimental results prove that the G-LSTM model has better prediction effect and overall performance compared with the selected traditional cross-project defect prediction methods.(3)In this paper,a small cross-project defect prediction visualization system is designed and implemented,which can demonstrate the process of G-LSTM crossproject defect prediction model and evaluate the prediction performance using three evaluation metrics,namely F1 metric,AUC and MCC.