Statistical analysis of three new measures of relevance redundancy and complementarity

Discriminant analysis is part of statistical learning; its goal is to separate classes defined a priori on a population and involves predicting the class of given data points. Discriminant analysis is applied in various fields such as pattern recognition, DNA microarray etc. In recent years, the discrimination problem remains a challenging task that has received increasing attention, especially for high-dimensional data sets. Indeed, in such a case, the feature selection is necessary, which implies the use of criteria of relevance, redundancy and complementarity of explanatory variables. The aim of this paper is to present an analysis of three new criteria proposed in this sense, more precisely based on the Principal Component Analysis we have been able to achieve a double objective: that of studying the harmony of these three criteria and also visualizing the class of candidate variables for a more in-depth selection in addition to eliminating the noise variables in a discriminant model.