The search for more optimal input spaces

Abstract

Designers of classifiers are faced with the problem of deciding which features should be used when building classifiers. The notion that adding extra features will always improve a classifier has been proved to be incorrect in the past. Thus, it is necessary to also investigate subsets of the full extracted feature set, to see whether better classification would not result. This feature input reduction also has an effect on cost and speed. Three methods for doing this input reduction are evaluated and compared. The methods yield encouraging results on real data sets. It is found that the gamma test method also has high correlation with classifier error rates, which might have a high impact on stopping criteria for neural networks.