Partitioning the feature space of a classifier with linear hyperplanes

Abstract

A method of constructing a linear hyperplane that partitions a multidimensional feature space with the objective of maximizing the mutual information associated with the partitioning is described. In addition, a process of constructing a decision-tree to hierarchically partition the training data using such hyperplanes is also introduced. The decision tree is used to quantize the feature space into nonoverlapping regions that are bounded by hyperplanes. The quantizer is also applied in conjunction with a Gaussian classifier in a speech recognition problem. Finally, the performance of this quantizer is compared with that of commonly used Gaussian clustering schemes.