摘要
近年来,软件缺陷预测逐渐成为软件工程领域的重要内容。很多典型的机器学习方法已经被应用到软件缺陷预测中,包括SVM、随机森林、决策树和朴素贝叶斯等。早期的研究工作对所有软件产品采取相同的特征提取方式,分类效果并不理想。后来一些特征选择方法被提出,比如基于启发试的回归特征消除方法已经成功与SVM方法结合起来,取得了较好的效果。文中在现有工作基础上借鉴了RFE(回归特征消除)的思想,考虑到朴素贝叶斯方法在处理小样本分类问题时的优越性,首次将RFE与朴素贝叶斯方法结合起来,利用贝叶斯模型的特性在特征选择后将特征权值应用到对分类决策的改进中,进一步提高了分类器性能。实验采用NASA的软件缺陷数据集,并对比了其他效果较好的分类算法,体现了该算法的优越性和有效性。
In recent years, software defect prediction is becoming an important part of the software engineering field. Many typical meth- ods like SVM, random forest, decision trees and Bayes have been applied to software defect prediction. However, earlier research almost takes the same feature set to train all kinds of software products and does not achieve a desired effect. Years later, some feature selection method are proposed. For example, the method combined recursive feature elimination and SVM, has got a good effect. In this paper, based on existing work, propose an algorithm which combines recursive feature elimination and Native Bayes. This algorithm will do a se- lection of feature set before training the model according to the contribution of each feature to get the optimal feature subset to be the in- put to train the model. The experiment adopts the software defect data set of NASA. Make a comparison with other machine learning methods, the experimental results demonstrate the superiority and effectiveness of this method.
出处
《计算机技术与发展》
2015年第10期131-134,139,共5页
Computer Technology and Development
基金
国家自然科学基金资助项目(61272273)
江苏省333工程项目(BRA2011175)
南京邮电大学校科研项目(XJKY14016)
关键词
软件缺陷
特征选择
朴素贝叶斯
缺陷预测
software defect
feature selection
Naive Bayes
defect prediction