Simple and interpretable discrimination

A number of approaches have been proposed for constructing alternatives to principal components that are more easily interpretable, while still explaining considerable part of the data variability. One such approach is employed in order to produce interpretable canonical variates and explore their discrimination behavior, which is more complicated as orthogonality with respect to the within-groups sums-of-squares matrix is involved. The proposed simple and interpretable canonical variates are an optimal choice between good and sparse approximation to the original ones, rather than identifying the variables that dominate the discrimination. The numerical algorithms require low computational cost, and are illustrated on the Fisher’s iris data and on moderately large real data.