On the problem of dimensionality and sample size in multi-stage pattern classifiers

In practical pattern recognition problems, the underlying probability distributions are not known a priori, but have to be estimated using finite number of labelled samples. It is well known that under such situations the Bayes classifier has a degrading performance when the number of features exceeds an optimal value. In this paper we study the possibility of using different classification procedures which use a subset of the available features at a step in an effort to circumvent the dimensionality problem. The classification schemes studied are the majority decision scheme and the decision tree classifier for normal populations.