A two-stage approach to feature downselection for pattern recognition

Abstract

Pattern recognition and classification algorithms have often relied upon the designer to determine which features are likely to be able to provide good discrimination, and given this set of features and a fixed model architecture to find the parameters that minimize a performance criterion. However, correlations among features can have damaging effects on the determination of decision boundaries in classifiers. Linear correlations can be identified easily, but nonlinear correlations, while also having detrimental effects, are usually not identified. By removing these, models can be made simpler and more robust. Additionally, other features may not have an appreciable effect on classifier performance (or even have a detrimental effect), and should be removed from the set of input features. By combining these two feature pruning approaches, input features can be substantially reduced without hurting classifier performance (and often helping). A data set of three-degree-of-freedom ballistics is used to demonstrate correlation removal and feature pruning.