In this paper the problems of a filtration of objects for metric algorithms of classification in particular, for an algorithm of the K nearest neighbors is considered. It is invited to change the class of anomalous objects which has similar regularities to improve the stability. First result and second result which was taken after the preprocessing are compared.