DISCRIMINANT-ANALYSIS USING NONMETRIC MULTIDIMENSIONAL-SCALING

Many discriminant methods produce a mapping of observations into the real line, and then an observation of an unknown group is allocated to a group according to its mapped position on the real line. In this paper the process is reversed so that training observations from each group are positioned in a Euclidean space, usually two-dimensional, using non-metric multidimensional scaling and then a mapping from the original sample space to the MDS space is found. This mapping is then used to discriminate future observations.