scientific article; zbMATH DE number 3905665

From MaRDI portal
Publication:3683358





In the classification problem where a large number of variables is involved, the probability of misclassification will be increased due to the accumulation of estimation errors. In order to reduce this probability, we have to choose the variables which have the largest contribution to the distance between the two distributions. But as soon as this selection is based upon estimations from samples, we are exposed to another danger, selecting the wrong variables which have large deviations of estimations from the corresponding mathematical expectations. In this circumstance, even the variables with zero contribution to the distance between two distributions may be selected; therefore, the probability of misclassification is increased. This effect has been analysed in some special cases by other authors. This paper deals with the general situation. The criterion \(J^ i\) and its estimation \(\hat J^ i\) represent the contribution of the variable \(X^ i\) to the Mahalanobis distance between two distributions and are used as the basis of selection of variables. Some mathematical relations between the quantities involved in the discriminant analysis are obtained. Among other facts, the existence of a limit value of the selection threshold is established and several important inequalities are derived. Three special cases are discussed. The influence of variables on the probability of misclassification for which the estimators \(\hat J^ 2\) deviate greatly from the expectations, is estimated. This topic is important for discriminant analysis both in theory and in application.











This page was built for publication:

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3683358)