Harmonic Averaging in Classifier Quality Assessment

The paper investigates quality measures of classifiers constructed on the basis of harmonic averaging. For a two-class (binary) classification, the measure P₄ is considered, which, in turn, is the harmonic mean of the precision and recall values for the cases where the first class is called positive, and, conversely, when the second class is called positive. New measures for assessing the quality of multiclass classifiers are proposed. As for the original measure F, we adhere to the idea of harmonic averaging and propose to use the harmonic mean of binary classifiers in the “One-Vs-Rest” scheme. The proposed measures have a number of unique properties from the point of view of classification problems: they allow taking into account the imbalance of the dataset and do not give preference to any of the classes. In addition, the proposed measures have good interpretability and are easy to calculate.