Discrimination and Classification model from Multivariate Exponential Power Distribution

Abstract

It is common to assume a normal distribution when discriminating and classifying a multivariate data based on some attributes. But when such data is lighter or heavier in both tails than the normal distribution, then the  probability of misclassification becomes higher giving unreliable result. This study proposed multivariate exponential power distribution a family of elliptically contoured model as underlining model for discrimination and classification. The distribution has a shape parameter which regulate the tail of the symmetric distribution to mitigate the problem of both lighter and heavier tails data, this generalizes the normal distribution and thus will definitely gives a lower misclassification error in discrimination and classification. The resulting discriminant model was compared with fisher linear discriminant function when applying to real data.