Application of the Mamdani fuzzy logic algorithm to assess the quality of artificial intelligence models based on available data

Abstract

This paper considers the actual problem of laboriousness and high cost of developing systems or models of artificial intelligence due to the low quality of the data used. The types and directions of the correlation between various metrics of data quality and the accuracy of the final artificial intelligence model are studied. It is noted that in the modern world there is a penetration of artificial intelligence into almost all areas of human activity: healthcare, agri-food, industry, creative areas. The relevance of the problem posed is substantiated and a brief review of modern studies related to the study of the correlation between data quality and artificial intelligence accuracy is carried out. It is noted that bad data lead to significant financial losses, increases the complexity of developing artificial intelligence systems or models. Based on the research conducted in various areas of artificial intelligence application, five characteristics of big data affecting the accuracy of the developed artificial intelligence have been identified: inconsistency, incompleteness, invalidity, noisiness, sample size. To determine the desired dependencies, the Mamdani fuzzy logic algorithm was chosen. Criteria for assessing data quality are converted into terms with fuzzy triangular numbers and fuzzy inference rules are formed. Dependence graphs are constructed and conclusions are drawn about the most important data quality criteria. High noise or data inconsistency are acceptable only in small quantities, but the accuracy of models drops sharply when these characteristics are enhanced, incompleteness or invalidity are less critical for the quality of models, increasing the sample size has a significant impact either with high model complexity or with a relatively small initial sample.