Generalized discriminant analysis for tumor classification with gene expression data

DNA microarray technology is the latest and the most advanced tool for parallel measuring of the activity and interactions of thousands of genes. The challenge is that the data dimension is large compared to the number of data points, which leads to small sample size (SSS) problem. Principal Component Analysis plus Linear Discriminant Analysis (PCA+LDA) is a well-known technique to cope with this problem, however, it cannot completely solve the SSS problem. In this paper we propose two novel discriminant techniques. Experimental results on gene expression data sets demonstrate that our methods have good discriminating power and outperform the direct linear discriminant analysis, moreover they are more stable than the PCA+LDA approach. © 2006 IEEE.