Evaluation of Classification Results

Abstrakt

The paper focuses on measures which can be applied to evaluation of classification results. Objects with known assignment to certain groups are used for evaluation. Using a classification method the user obtains the assignment of objects to groups. Many coefficients have been proposed for evaluation of the success rate of classification. Most of them is determined for classification to two groups. Possibilities for classification to more groups are limited. The aim of this paper is to summarize different measures, discuss their origin and relationships. For classification to three or more groups we propose two novel measures which are more suitable. The first of them takes a variability of diagonal frequencies of the confusion matrix into account. The second one is based on the sum of squared differences between the maximum correctly assigned objects and real correctly assigned objects in each group.