Evolutionary inspired optimization of feature subset ensembles

We propose a framework for searching for the sets of feature subsets that can serve as the bases for learning local classifiers within a classifier ensemble. We follow the methodology based on evaluation of random feature permutations that produce locally optimal feature subsets. We investigate various heuristics for selection of feature subsets (permutations) into the ensembles, as well as for evaluation of those ensembles with respect to their predicted classification accuracy and their ability to clearly represent dependencies in data. The obtained framework is verified on four representative benchmark data sets. Further research directions are identified.