A1663
Title: Theoretical foundations and asymptotic confidence intervals for multiclass Matthews correlation coefficients
Authors: Jun Tamura - Yokohama City University (Japan) [presenting]
Kouji Yamamoto - Yokohama City University (Japan)
Kenichi Hayashi - Keio University (Japan)
Abstract: Classification problems arise in many fields and require appropriate performance measures. In binary classification, the Matthews correlation coefficient (MCC) has attracted considerable attention because it incorporates all cells of the confusion matrix and provides a more comprehensive summary than some commonly used alternatives. In many practical applications, however, classification problems involve more than two categories, and several extensions of MCC have been proposed for multiclass settings. Although these measures are increasingly used, their theoretical relationships have not yet been fully clarified, and methods for statistical inference remain limited. The theoretical structure of MCC-based measures for multiclass classification is clarified and asymptotic confidence intervals for these measures are derived. Paired designs, in which multiple classifiers are applied to the same set of subjects, are also considered, and asymptotic confidence intervals for differences between MCC-based measures are constructed. Simulation studies are conducted to evaluate the finite-sample performance of the proposed procedures and related methods.
