基于迁移学习的跨项目软件缺陷预测  

Cross-project Software Defect Prediction Based on Transfer Learning

作　　者：张洋洋 荆晓远[1] 吴飞[1] ZHANG Yang-yang;JING Xiao-yuan;WU Fei(School of Automation,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学自动化学院,江苏南京210003

出　　处：《计算机技术与发展》2018年第12期83-85,90,共4页Computer Technology and Development

基　　金：国家自然科学基金(61702280);江苏省自然科学基金(BK20170900);江苏省高等学校自然研究项目(17KJB520025);南京邮电大学引进人才科研启动基金(NY217009)

摘　　要：软件缺陷预测在提高软件质量、控制与平衡软件成本方面起着举足轻重的作用,是软件工程的活跃领域。研究者们提出了许多预测技术,从不同层面解决了不同的问题。传统软件缺陷预测算法在面对跨项目软件缺陷预测中往往不能得到一个好的结果,原因是训练数据样本(源数据)和测试数据样本(目标数据)之间的分布是不同的。为了解决这个问题,提出了一种基于迁移学习的跨项目软件缺陷预测算法。该算法首先采用了一种不同分布之间的距离度量方式,训练出一种模型来最小化训练数据和测试数据之间的分布差异以及条件分布差异,在映射过后的新的特征空间中两种数据集几乎拥有同样的分布。然后就可以采用传统的机器学习算法进行分类。实验结果表明,该算法具有较好的预测性能。Software defect prediction plays an important role in improving software quality,controlling and balancing software costs as an active area of software engineering.Researchers have proposed many prediction techniques to solve different problems at different levels.The traditional software defect prediction algorithm often cannot get a ideal results in the face of cross-project software defect prediction because the distribution between the training data sample(source data)and the test data sample(target data)is different.In order to solve this problem,we present a cross-project software defect prediction algorithm based on transfer learning.Firstly,the algorithm uses a distance measurement between different distributions to train a model to minimize the distribution difference and conditional distribution difference between training data and test data.After mapping the two data sets have almost the same distribution in the new feature space.Then the traditional machine learning algorithm can be used for classification.The experiment shows that the proposed algorithm has better predictive performance.

关 键 词：软件缺陷预测 迁移学习 特征映射 机器学习 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]