Class-specific Linear Discriminant Analysis
Songcan Chen · Journal of Chinese Computer Systems · 2008
Linear discriminant analysis(LDA) has been successfully used as a feature extraction method.LDA aims to project the samples onto a feature space,where the ratio of the between-class scatter to the within-class scatter of the projected samples is maximized so the projected samples are well separated.However,this method projects all the samples to the same feature space with the same criteria and the difference between the distributions of the data of different classes is ignored.In this paper a new class-specific linear discriminant analysis method is proposed,which searches for the optimal transformation matrix for each class,to ensure this class samples are well separated from all the other samples in the projected space.And then obtain class-specific linear discriminant analysis method in the empirical feature space by combining the method with empirical kernel.The experimental results on artificial dataset and UCI benchmark datasets show that CSLDA outperforms LDA both in the input space and the empirical feature space.