On the choice of partial orders on feature values sets in the supervised classification problem

The authors consider one of the central problems of machine learning — the supervised classification. A scheme for the logical classification algorithms synthesis is described under the assumption that the features descriptions of precedents are the elements of the finite partial orders Cartesian product. A criterion for the correctness of the voting algorithm of representative elementary classifiers is formulated. The authors study the possibility of defining linear orders on sets of feature values that provide better classification, which is not necessarily correct, in assumption that the source data are not ordered (the precedents descriptions are the elements of the antichains product). A procedure is proposed for “correct” consistent ordering of the acceptable values of separate features, while the remaining features are antichains. The results of experiments on real data are presented demonstrating the effectiveness of the methods developed in the work.