Selection of informative features based on discrete-fractional functionals

Аннотация

Currently, the most common criteria for informativeness of features are heuristic criteria associated with evaluating the separability of specified classes and based on the compactness hypothesis that is fundamental to pattern recognition: with increasing distance between classes, their separability improves. "Good"are those signs that maximize this distance. Such heuristic criteria, although they are widely used in solving practical problems of classification, however, in theoretical terms, are little studied. At present, the method of selection of informative features with regard to interrelations of features based on heuristic criteria has not been developed. The article deals with the problem of choosing informative features based on heuristic criteria, taking into account their interrelationships. The optimal conditions for the “ordering” method and the selected vector are also determined. Using proven theorems, a new method for selecting informative features using discrete-fractional functionals has been developed.