关键词: 典型相关分析, 线性判别分析, 特征提取, 类依赖, 多类分类器设计

Abstract: Canonical Correlation Analysis（CCA） is an important dimension reduction method for feature extraction by virtue of the correlation between samples.Correlation Discriminant Analysis（CDA） maximizes the correlation of the same class samples and minimizes the correlation between different class samples in the feature space，which can be treated as class-specific CCA.The feature extraction of the two methods is independent of the following classifier，i.e.the extracted features unnecessarily favor the classification.This independence affects the performance of classifiers unavoidably.In virtue of the characteristics that the vertices of the regular simplex are equidistant and affine invariant，use the vertices of the regular simplex to code the class labels so that the class information is incorporated in the classifier design.The algorithm，termed as Class-specific Correlation Multi-class Classifier（CCMC），maximizes the correlation between each sample and its class label，minimizes the correlation between each sample and the others class labels at the same time.Furthermore，this paper also proposes the corresponding kernelized version of CCMC by combining with the Empirical Kernel mapping，called as EK-CCMC.The experimental results on both artificial dataset and UCI benchmark datasets show that the proposed CCMC and EK-CCMC enhance the classifier performance by using class-specific correlation to design classifier.

Key words: Canonical Correlation Analysis（CCA）, linear discriminant analysis, feature extraction, class-specific, multi-class classifier design