«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

27

期数:

2017年07期

页码:

38-42

栏目:

智能、算法、系统工程

出版日期:

2017-07-10

文章信息/Info

Title:

 Defect Prediction of Co-training Software with Non-negative Sparse Graph

文章编号:

1673-629X（2017)07-0038-05

作者:

 张志武[1] ; 荆晓远[2][3];  吴飞[2]

 1.南京邮电大学 计算机学院；2.南京邮电大学 自动化学院；3.武汉大学 软件工程国家重点实验室

Author(s):

 ZHANG Zhi-wu[1] ; JING Xiao-yuan[2] [3]; WU Fei[2]

关键词:

 ; 非负稀疏图; 协同训练; 半监督学习; 软件缺陷预测

Keywords:

 ; non-negative sparse graph; co-training; semi-supervised learning; software defect prediction

分类号:

TP311

文献标志码:

A

摘要:

 软件缺陷预测是一种可提高软件系统质量和优化测试资源分配的软件系统可靠性保证方法.当软件历史仓库中有标记训练模块较少时,应用机器学习方法构建有效的预测分类器是一个有挑战性的问题.为此,提出了一种基于非负稀疏图的协同训练软件缺陷预测方法,该方法汇集基于图的半监督学习方法和协同训练方法的优点,对无标记数据进行显示置信度估计.其利用软件模块间的相似性构建一个非负稀疏图,图中边的权重反映了样本间的相似度;利用协同训练的三个分类器对无标记样本的隐式选择和显示计算其所属类别的置信度,选取可靠的无标记样本辅助有标记样本进行训练以减少噪声数据的引入,并逐个迭代更新分类器,直至达到最大迭代次数或分类器识别率降低为止.基于NASA MDP数据集的验证实验结果表明,所提出的方法优于具有代表性的半监督协同训练方法.

Abstract:

 Software defect prediction is a system reliability assurance method which can improve the quality of software system and optimize the distribution of test resources.When the previous defect labels of modules in software history warehouse are limited,building an effective prediction classifier by using machine learning methods becomes a challenging problem.Aiming at this problem,a co-training algorithm for software defect prediction based on non-negative sparse graph is proposed,which combines with the advantages of the graph-based semi-supervised learning method and the co-training method and estimates the confidence of unlabeled data.A non-negative sparse graph has been constructed by the similarity between the software modules so that the edge of the graph reflects the similarity between samples.Then three classifiers have been employed for co-training.In order to reduce the introduction of noise data,the reliable unlabeled samples have been selected for training by the implicit selection of the three classifiers and the confidence estimation of the categories.The classifiers keep to iteratively updating until the maximum number of iterations has reached or the recognition rates of classifiers have been reduced.Experimental results on NASA MDP datasets show that the proposed method is superior to the representative semi-supervised co-training method.

常用功能

更新日期/Last Update: 2017-08-22