一种面向软件缺陷预测的可容忍噪声的特征选择框架

发布时间：2019-05-27 06:14

【摘要】：软件缺陷预测通过挖掘软件历史仓库,构建缺陷预测模型来预测出被测项目内的潜在缺陷程序模块.但在挖掘过程中,对程序模块进行类型标记或软件度量时均可能产生噪声.虽然研究人员对已有特征选择方法的噪声容忍能力进行了分析,但据我们所知,很少有研究人员在软件缺陷预测研究中,针对性的设计出可容忍噪声的新颖特征选择方法.为了解决此问题,我们提出一种可容忍噪声的特征选择框架FECS.具体来说,首先借助聚类分析,将原始特征集划分到指定数目的簇中,随后设计出3种不同的启发式特征选择策略,依次从每一个簇中选出最为典型的特征.在实证研究中,以Eclipse和NASA等实际项目为评测对象.首先借助一系列数据预处理方法来提升数据集质量,随后同时注入类标噪声和特征噪声来模拟噪声数据集.通过与典型的特征选择方法进行比较,验证了FECS框架的有效性,除此之外,通过深入分析噪声注入率、特征选择比例及噪声类型对缺陷预测性能的影响,为更有效的使用FECS提供了指导.
[Abstract]:Software defect prediction by mining software history warehouse and constructing defect prediction model to predict the potential defect program module in the project under test. However, in the process of mining, noise may be generated when the program module is marked by type or measured by software. Although the researchers have analyzed the noise tolerance of the existing feature selection methods, as far as we know, few researchers have designed a novel feature selection method which can tolerate noise in the software defect prediction research. To solve this problem, we propose a noise tolerant feature selection framework (FECS.). Specifically, with the help of cluster analysis, the original feature set is divided into a specified number of clusters, and then three different heuristic feature selection strategies are designed, and the most typical features are selected from each cluster in turn. In the empirical study, the actual projects such as Eclipse and NASA are taken as the evaluation objects. Firstly, a series of data preprocessing methods are used to improve the quality of the data set, and then the quasi-standard noise and feature noise are injected at the same time to simulate the noise data set. The effectiveness of FECS framework is verified by comparing with the typical feature selection method. In addition, the effects of noise injection rate, feature selection ratio and noise type on defect prediction performance are deeply analyzed. It provides guidance for more effective use of FECS.
【作者单位】：南京大学计算机软件新技术国家重点实验室;南京大学计算机科学与技术系;南通大学计算机科学与技术学院;
【基金】：国家自然科学基金(61373012,61202006,61321491,91218302) 欧盟项目(612212) 南京大学计算机软件新技术国家重点实验室开放课题(KFKT2016B18);南京大学优秀博士研究生创新能力提升计划和软件新技术与产业化协同创新中心部分资助
【分类号】：TP311.5

【相似文献】