A heuristic for feature selection for the classification with neural nets

Feature selection is a very important task in classification, but it is an NP-hard optimization problem that is often approximated by heuristic search for a good feature subset. To compare different feature subsets according to their relevance for a classification problem two approaches are used. Filters compute scores on general characteristics of the data without reference to a classifier. Wrappers generate a set of features, build a classifier using these features and use the classification rate to evaluate the feature set. Wrapper models are computationally more expensive than filter models, but for a given classifier the result of a wrapper is more reliable. The idea in this work is to use a wrapper with a fast statistical classifier for feature selection, while for the later online classification in the application a neural net is used. Some requirements must be met to ensure a correspondence between the classifiers such that the feature subset evaluated by the statistical classifier correlates to the features needed by the neural net for the differentiation of classes. Combined with different search methods this approach was used to build a small classifier for acoustic signals that can be implemented in hearing aids.