改进加权朴素贝叶斯的软件缺陷预测算法  

A Software Defect Prediction Algorithm Using Improved Weighted Naive Bayesian

作　　者：郭树强[1] 黄蕊[2] 李卿 GUO Shu-qiang;HUANG Rui;LI Qing(Department of Information Technology,Hebei Vocational College for Correctional Police,Shijiazhuang 050081 China;Instiute of Arifcial Ielligence,Chengde University of Radio and Television,Chengde 067000,China;Polytechnic College,Hebei University of Science and Technology,Shijiazhuang 050041,China)

机构地区：[1]河北司法警官职业学院信息技术部,河北石家庄050081 [2]承德广播电视大学人工智能研究院,河北承德067000 [3]河北科技大学理工学院,河北石家庄050041

出　　处：《控制工程》2021年第3期600-605,共6页Control Engineering of China

基　　金：河北省高等学校科学技术研究项目(z2015167)。

摘　　要：针对软件缺陷预测领域特征之间存在紧密关联性而影响朴素贝叶斯分类性能的问题,提出一种改进加权朴素贝叶斯的分类算法。首先,通过预处理步骤以及特征成对计算,创建彼此之间的依赖关系。然后,通过构造权值的方式实现朴素贝叶斯的独立性假设松弛。最后,使用离散化方法将软件指标的数值转化为分类值,并利用min-max归一化程序对数据进行归一化处理。使用获得广泛认可的NASA信用数据库进行实验,实验结果表明,提出的改进算法比标准朴素贝叶斯方法预测效果更好,相比其他几种同类算法,提出的算法更具竞争性。For the issue that the close correlation between features in software defect prediction applications affects the performance of naive Bayesian classification, an improved weighted naive Bayesian classification algorithm is proposed. Firstly, the dependencies between them are created by pre-processing steps and pairwise computation of features. Then, the independence hypothesis of naive Bayesian is relaxed by constructing weights. Finally, the value of software index is transformed into the classification value by discretization method, and the data is normalized by min-max normalization program. The experiments are carried out by using widely recognized NASA promise data sets. The results show that the proposed algorithm has better prediction effect than the standard naive Bayesian method, and it is more competitive than other feature-weighted techniques.

关 键 词：加权朴素贝叶斯 特征不相关 加权系数 软件缺陷预测 离散化 min-max归一化 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]