Multivariate statistical pattern recognition with nonreduced dimensionality
Jǐŕı Grim · Czech digital mathematics library · 1986
WITH NONREDUCED DIMENSIONALITYA direct solution of multivariate classification problems is considered without preceding reduction of dimensionality.The unknown class-conditional distributions are approximated by finite mixtures of special type.For decision purposes the components of these mixtures can be reduced to functions defined on different subspaces.To optimize the choice of subspaces and of the related parameters maximum-likelihood principle is used.The corresponding m.-l.estimates are computed by the EM algorithm.In this way the feature selection problem can be solved independently for each component of the approximating mixtures without introducing any additional criteria.