【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。更多还原

【Abstract】 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.更多还原