Comparison of Different Dichotomous Classification Algorithms
Yu. I. Zhuravlev, V. V. Ryazanov, Vl. V. Ryazanov, Levon Aslanyan, Hasmik Sahakyan
Abstract
Experimental investigations of various dichotomous classification algorithms are carried out. Dichotomous classification, or Error-Correcting Output Codes (ECOCs) classification, is based on the construction of a binary code matrix. The rows of the matrix contain unique codewords of classes, and columns are called dichotomies. A dichotomous classification consists of two stages: coding (construction of a code matrix) and decoding, making a decision on the correspondence of an object to a class by analyzing the code matrix. In this study, an experimental comparison of newly proposed methods for constructing dichotomies and a comparison of different approaches to decoding by the available code matrix are proposed. Preliminary experiments show the prospects of proposed methods.