A CLASS DISCRIMINABILITY MEASURE BASED ON FEATURE SPACE PARTITIONING

摘要

This paper presents a new class discriminability measure based on an adaptive partitioning of the feature space according to the available class samples. It is intended to be used as a criterion in a classifier-independent feature selection procedure. The partitioning is performed according to a binary splitting rule and appropriate stopping criteria. Results from several tests with Gaussian and non-Gaussian, multidimensional and multiclass computer-generated samples, were very similar to those obtained using a Bayes error criterion function, i.e. the optimal feature subsets selected by both criterion functions were the same. The main advantage of the new measure is that it is computationally efficient. [References: 17]