Comparative Study of Quality Estimation of Binary Classification

The paper describes results of analytical and experimental analysis of seventeen functions used for evaluation of binary classification results of arbitrary data. The results are presented by 2×2 error matrices. The behavior and properties of the main functions calculated by the elements of such matrices are studied. Classification options with balanced and imbalanced datasets are analyzed. It is shown that there are linear dependencies between some functions, many functions are invariant to the transposition of the error matrix, which allows us to calculate the estimation without specifying the order in which their elements were written to the matrices.It has been proven that all classical measures such as Sensitivity, Specificity, Precision, Accuracy, F1, F2, GM, the Jacquard index are sensitive to the imbalance of classified data and distort estimation of smaller class objects classification errors. Sensitivity to imbalance is found in the Matthews correlation coefficient and Kohen’s kappa. It has been experimentally shown that functions such as the confusion entropy, the discriminatory power, and the diagnostic odds ratio should not be used for analysis of binary classification of imbalanced datasets. The last two functions are invariant to the imbalance of classified data, but poorly evaluate results with approximately equal common percentage of classification errors in two classes.We proved that the area under the ROC curve (AUC) and the Yuden index calculated from the binary classification confusion matrix are linearly dependent and are the best estimation functions of both balanced and imbalanced datasets.