Discrimination Using Covariance Selection Models in the Automated Allocation of Human Chromosomes

Abstract

Various methods have been proposed for modifying estimates of covariance matrices in normal-based discriminant analysis to reduce the effect of sampling variability on discriminant functions. We consider the use of covariance selection models for this purpose in the context of quadratic discrimination for the automated allocation of human chromosomes to 24 classes. We investigate the use of simple method for deciding which partial correlations should be set to 0 and show that the resulting discriminant functions can both reduce error rates and speed computation when the number of features is large.