DISCRIMINANT-ANALYSIS BY GAUSSIAN MIXTURES

Fisher-Rao linear discriminant analysis (LDA) is a valuable tool for m ultigroup classification. LDA is equivalent to maximum likelihood clas sification assuming Gaussian distributions for each class. In this pap er, we fit Gaussian mixtures to each class to facilitate effective cla ssification in non-normal settings, especially when the classes are cl ustered. Low dimensional views are an important by-product of LDA - ou r new techniques inherit this feature. We can control the within-class spread of the subclass centres relative to the between-class spread. Our technique for fitting these models permits a natural blend with no nparametric versions of LDA.