基于核主成分分析的超多面体数据描述方法  

Polytope combination data description method based on kernel principal component analysis

作　　者：陈宇晨 何毅斌[1] 戴乔森 刘湘 CHEN Yu-chen;HE Yi-bin;DAI Qiao-sen;LIU Xiang(School of Mechanical and Electrical Engineering,Wuhan Institute of Technology,Wuhan 430205,China)

机构地区：[1]武汉工程大学机电工程学院,湖北武汉430205

出　　处：《计算机工程与设计》2021年第3期663-669,共7页Computer Engineering and Design

基　　金：湖北省科技厅重大专项基金项目(2016AAA056);化工装备强化与本质安全湖北省重点实验室开放基金项目(2018KA01);武汉工程大学研究生教育创新基金项目(CX2019243)。

摘　　要：为更加准确地从大量数据中检测出存在的小量异常数据,提出基于核主成分分析的超多面体数据描述方法。利用核主成分分析对数据进行非线性映射,在此基础上使用主成分信息在N维空间中建立2N个超平面,组合这些超平面构造一个超多面体模型,为现有的数据描述类方法提供更多选择。通过第三方数据集并与支持向量数据描述方法比较,验证了在某些分布下超多面体比超球体的支撑域适合训练数据,得到了更好的分类效果,表明了该方法的有效性。To more accurately detect a small amount of abnormal data from a large amount of data,a combination data description method based on kernel principal component analysis was proposed.Kernel principal component analysis was used to carry out nonlinear mapping on the data,on this basis,2N hyperplanes were established in N-dimensional space using principal component information,these hyperplanes were combined to construct a combination data description model,thus providing more choices for the existing data description class methods.Through a third-party data set and comparison with a support vector data description,it is verified that the hyper-polyhedron is more suitable for training data than the support domain of the hyper-sphere under some distributions,a better classification result is obtained,which shows the effectiveness of the method.

关 键 词：异常数据 核主成分分析 超多面体数据描述 支持向量数据描述 支撑域 

分 类 号：TP306.3[自动化与计算机技术—计算机系统结构]