融合数据分布特征的多视图典型相关分析  

Multi-view canonical correlation analysis with data distribution

作　　者：郭慧 刘忠宝[2] GUO Hui;LIU Zhong-bao(School of Information,Business College of Shanxi University,Taiyuan 030031,China;School of Software,North University of China,Taiyuan 030051,China)

机构地区：[1]山西大学商务学院信息学院,山西太原030031 [2]中北大学软件学院,山西太原030051

出　　处：《广西大学学报（自然科学版）》2018年第6期2252-2257,共6页Journal of Guangxi University（Natural Science Edition）

基　　金：山西省自然科学基金资助项目(201601D011042);山西省高等学校创新人才支持计划项目(2016)

摘　　要：典型相关分析CCA是一种经典的特征提取方法,该方法找到的投影方向满足两视图数据集之间的相关性最大,该方法在生产实践中广泛应用,但当面对多视图数据时便无能为力。鉴于此,研究人员提出多视图相关分析MCCA,MCCA为每个视图找到一组投影方向,并保证投影后的视图数据集之间的相关性最大,该方法有效地拓展了CCA的适用范围。但MCCA在特征提取时,并未考虑数据的分布性状,因而,其工作效率有待于进一步提升,基此提出融合数据分布特征的多视图典型相关分析MCCA-DD,该方法引入类间离散度和类内离散度,用以表征数据的分布性状,以期提高MCCA的特征提取效率。从学生体测成绩数据集和多特征手写体数据集上的比较实验表明:与典型相关分析CCA、多视图相关分析MCCA等特征提取方法相比,MCCA-DD具有更优的特征提取效率。Canonical Correlation Analysis(CCA)is a typical feature extraction method,which tries to find an optimal projection to maximize the correlation between the two-view datasets.CCA is widely used in practice,while it cannot deal with the multi-view datasets.In view of this,Multi-view Canonical Correlation Analysis(MCCA)is proposed by researchers.MCCA tries to find an optimal projection for each view,and verifies the correlation between the multi-view datasets maximized.However,MCCA does not take the data distribution into consideration during feature extraction,so the efficiency needs to be further improved.Multi-view Canonical Correlation Analysis with Data Distribution(MCCA-DD)is proposed in this paper.The between-class scatter and the within-class scatter are introduced into the method to describe the data distribution in order to improve the feature extraction efficiencies of MCCA.Comparative experiments on the student s sports test datasets and UCI datasets verify that MCCA-DD has better feature extraction efficiency than CCA and MCCA.

关 键 词：特征提取 多视图典型相关分析 数据分布特征 类间离散度 类内离散度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]