On the number of partial least squares components in dimension reduction for tumor classification

Abstract

Dimension reduction is important during the analysis of gene expression microarray data, because the high dimensionality of data sets hurts the generalization performance of classifiers. Partial Least Squares (PLS) based dimension reduction is a frequently used method, since it is specialized in handling high dimensional data set and leads to satisfying classification performance. This paper investigates the influence on generalization performance caused by the variation of the number of PLS components and the relationship between classification performance and regression quality of PLS on the training set. Experimental results show that the number of PLS components for classifiers can be automatically determined by regression quality of PLS latent variables. © Springer-Verlag Berlin Heidelberg 2007.