ПРАВДОПОДІБНІ, АЛЕ НЕОБГРУНТОВАНІ ПЕРЕДУМОВИ ПРИ ПОБУДОВІ ДІАГНОСТИЧНИХ МОДЕЛЕЙ

• Дата: 2020
• Автор: Fainzilberg, L.S.
• Формат: Стаття
• Мова: English
• Опубліковано: V.M. Glushkov Institute of Cybernetics of NAS of Ukraine 2020
• Теми: дискримінантна функція; діагностична ознака; методи самоорганізації
• Онлайн доступ: https://jais.net.ua/index.php/files/article/view/474
• Назва журналу: Problems of Control and Informatics

When solving a number of applied problems of medical and technical diagnostics, the construction of diagnostic models is carried out in conditions of insufficient knowledge of the physical laws that arise in the object of study. It is necessary to build models only on the basis of common sense and intuition, relying on the available experimental material (precedents). However, in this case, erroneous solutions are possible that lead to the inefficiency of the diagnostic system. The article discusses examples of some arguments that are not scientifically sound. It is shown that the linear classifier, to which the method of decision making on the distance to standards is reduced, can lead to absurd results. Such an effect occurs if the independence condition for the preferences of individual characteristics is not fulfilled, which means that the «deterioration» of the value of one attribute can be compensated by the «improvement» of the other, and vice versa, which is not always true. It is shown that unreasonable expansion of the space of diagnostic signs can worsen the effectiveness of the diagnostic rule. Therefore, it is important to get rid of unnecessary signs even before the training phase. Inconsistency of the argument that when constructing diagnostic models it is advisable to use only statistically independent attributes is analyzed. To illustrate the fallacy of such argument it is proved that with a statistical relationship between features, a combination of individual non-informative features can be not only useful, but also provide error-free recognition of classes. Therefore, it is important in each case to investigate the issue of conditional statistical dependence between features before making a decision on their exclusion from the description. Using the example of constructing a model for indirect estimation of the carbon content in a liquid metal by the temperature of crystallization onset, it is shown that it is impossible to restore the true diagnostic model using only self-organization methods without using additional algorithms.