https://arxiv.org/pdf/1905.05667

https://www.crcpress.com/Classification-2nd-Edition/Gordon/p/book/9781584880134

http://scikit-learn.org/stable/modules/clustering.html

http: // www- bcf.usc.edu/~gareth/ISL/ISLR%20Sixth%20Printing.pdf

http://arxiv.org/pdf/0909.3052.pdf

As anomaly detection algorithms aim to classify whether the target is an anomaly or not, it falls under binary classification. So, mostly the evaluation metrics used are accuracy, precision and recall and ROC-AUC curve. Usually, the number of anomaly samples will be less when compared to normal samples. Hence just accuracy measure is not a correct evaluation metric. The best and suitable evaluation metric for anomaly detection algorithms is the ROC-AUC curve.

INVARIANT TO VOLUMES OF DATA, A FUZZY MULTICLASS GENERALIZATION OF F-MEASURE OF PLAUSIBILITY IN VAN RIJSBERGEN MODELS IN ASC-ANALYSIS AND IN THE "EIDOS" SYSTEM

Classic quantitative measure of the reliability of the models: F-measure by van Rijsbergen is based on counting the total number of correctly and incorrectly classified and not classified objects in the training sample. In multiclass classification systems, the facility can simultaneously apply to multiple classes. Accordingly, when the synthesis of the model description is used for formation of generalized images of many of the classes it belongs to. When using the model for classification, it is determined by the degree of similarity or divergence of the object with all classes, and a true-positive decision may be the membership of the object to several classes. The result of this classification may be that the object is not just rightly or wrongly relates or does not relate to different classes, both in the classical F-measure, but rightly or wrongly relates or does not relate to them in varying degrees. However, the classic F-measure does not count the fact that the object may in fact simultaneously belongs to multiple classes (multicrossover) and the fact that the classification result can be obtained with a different degree of similarity-differences of object classes (blurring). In the numerical example, the author states that with true-positive and true-negative decisions, the module similarities-differences of the object classes are much higher than for false-positive and false-negative decisions. It would therefore be rational to the extent that the reliability of the model to take into account not just the fact of true or false positive or negative decisions, but also to take into account the degree of confidence of the classifier in these decisions. In classifying big data we have revealed a large number of false-positive decisions with a low level of similarity, which, however, in total, contribute to reducing the reliability of the model. To overcome this problem, we propose a L2- measure, in which instead of the sum of levels of similarity we use the average similarity by different classifications. Thus, this work offers measures of the reliability of the models, called L1-measure and the L2 measure, mitigating and overcoming the shortcomings of the F-measures; these measures are described mathematically and their application is demonstrated on a simple numerical example. In the intellectual system called "Eidos", which is a software toolkit for the automated system-cognitive analysis (ASC-analysis), we have implemented all these measures of the reliability of the models: F, L1 and L2

Lutsenko E. V. Invariant to volumes of data, a fuzzy multiclass generalization of F-measure of plausibility in van Rijsbergen models in ASC-analysis and in the "Eidos" system / E. V. Lutsenko / / Polytematic network electronic scientific journal of the Kuban state agrarian University (scientific journal of kubgau) [Electronic resource]. - Krasnodar: Kubgau, 2017. – No. 02(126). C. 1 – 32. – IDA [article ID]: 1261702001. - Access mode: http://ej.kubagro.ru/2017/02/pdf/01.pdf, 2 C. p. l.