流式数据的动态正交降维算法  

Dynamic Algorithm for Orthogonal Dimension Reduction on Streaming Data

作　　者：陈新元 谢晟祎[2] 陈庆强 刘羽 CHEN Xinyuan;XIE Shengyi;CHEN Qingqiang;LIU Yu(Department of Information Engineering, Fuzhou Melbourne Polytechnic, Fuzhou, Fujian 350108, China;Experimental Training Center, Fujian Vocational College of Agriculture, Fuzhou, Fujian 350181, China;Information Science and Engineering College, Fujian University of Technology, Fuzhou, Fujian 350118, China;Modern Education Technical Center, Fuzhou Melbourne Polytechnic, Fuzhou, Fujian 350108, China)

机构地区：[1]福州墨尔本理工职业学院信息工程系,福建福州350108 [2]福建农业职业技术学院实验实训中心,福建福州350181 [3]福建工程学院信息科学与工程学院,福建福州350118 [4]福州墨尔本理工职业学院现代教育技术中心,福建福州350108

出　　处：《闽江学院学报》2020年第5期47-57,共11页Journal of Minjiang University

基　　金：福建省中青年教师教育科研项目(JAT191663)。

摘　　要：针对行业数据增长快、维数高且核心特征不明确等特点,提出基于Schmidt正交分解的动态正交降维算法(DOADR),在线性独立性基础上,设计自适应阈值机制以筛选基底矩阵,同时实现维数动态增长;根据流式数据的特点进一步提出增量DOADR算法优化阈值设计。真实数据集上的实验证明,算法较主流模型在计算开销、重构误差和分类准确率等指标上具有一定优势。In view of characteristics of industry data,such as fast growth speed,multiple dimensions and unclear key features,an algorithm called dynamic orthogonal analysis for dimension reduction(DOADR)is proposed based on Schmidt decomposition.With the definition of linear independence,a self-adaptive threshold mechanism is designed to generate basement matrix and achieve dynamic growth of dimensions.Furthermore,in incremental DOADR,threshold optimization is performed taking the characteristics of streaming data into consideration.Experiments on real-world datasets prove that compared with mainstream models our algorithms hold certain advantages on computational cost,reconstruction error and classification accuracy.

关 键 词：降维 正交分解 动态阈值 流式数据 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]