Pairwise Selection of Features and Prototypes

Abstract

Learning from given patterns is realized by learning from their appropriate representations. This is usually practiced either by defining a set of features or by measuring proximities between pairs of objects. Both approaches are problem dependent and aim at the construction of some representation space, where discrimination functions can be defined. In most situations, some feature reduction or prototype selection is mandatory. In this paper, a pairwise selection for creating a suitable representation space is proposed. To determine an informative set of features (or prototypes), the correlations between feature pairs are taken into account. By this, some dependencies are detected, while overtraining is avoided as the criterion is evaluated in two-dimensional feature spaces. Several experiments show that for small sample size problems, the proposed algorithm can outperform traditional selection methods.