Selecting relevant projections onto subsets of coordinates: A minimax dependence-based approach (Q750027)

Given a set of entities and a random vector whose components are defined on that set, the sum of entropies of the components minus the joint entropy (or entropy of the random vector) is used to reduce the number of components. This reduction is accomplished by selecting a pair (or triple) of components that are significant for classifying the set of entities in distinct and meaningful clusters. This purpose is achieved by performing the selection on the basis of both, the low degree of interdependence among selected components, and the high degree of dependence on the omitted ones (that means a kind of minimax criterion). When components are Gaussian, the measure of interdependence admits a simple expression that can be easily computed through MINITAB or SAS procedures. Results are illustrated by means of a practical example.