Criteria for training fuzzy classifier taking into account the cost matrix

Abstract

The tie "input – output" is described by linguistic if – then rules where antecedents contain fuzzy terms "low", "medium", "high" in the fuzzy classifiers. To enhance the correctness it is necessary to train fuzzy classifier on experimental data. The paper extends the criteria for training fuzzy classifier to the case of the cost matrix, which consist of costs of the different types of errors. Computer experiments on the task of heart disease diagnosis have shown that the best quality setting enables the use of the training criterion, in which the distance between the fuzzy inference results and experimental data for the cases of misclassification is multiplied by penalty coefficient.