Decision analysis

Feature selection method based on a probabilistic approach and cross-entropy metric for image recognition problem

Yu. A. Dubnov<sup>ab

a</sup> Federal Research Center "Computer Science and Control" of Russian Academy of Sciences, Moscow, Russia
^b HSE University, Moscow, Russia

Abstract: The paper considers the problem of feature selection in the classification problem. A method for selecting informative features based on a probabilistic approach and cross-entropy metrics is proposed. Several variants of the information criterion for selecting features for a binary classification problem are considered, as well as its generalization to the case of a multiclass problem. Demonstration examples of the proposed method for the task of image recognition from the mnist collection are given.

Keywords: feature selection, classification, cross-entropy.

DOI: 10.14357/20718594200206

Bibliographic databases:

•