Increasing quality of classifying objects using new metrics of clustering

The article touches upon one of the main problems of machine learning — clustering objects. It has been widely used in various subject areas: marketing, sociology, psychology, etc. Clusterization algorithms, as a rule, are based on a metric that reflects the distance between objects. However, in some cases it is not practical to use the distance between objects. In certain situations, it is possible to say that one object is similar to the other, the latter being not similar to the former. The original picture and its copy may serve as an example. For such cases, a measure of object similarity is proposed in the work, which shows how many features of one object are contained in another one. A similarity matrix is built on this measure, the analysis of which allows revealing clusters of mutually similar objects. When testing the proposed clustering method, the Rand index (the proportion of correctly connected or unrelated objects) made 0.93. There has been proposed an algorithm that allows to form a set of objects absolutely different from each other. A set of objects formed in this way can later become a learning set for classifiers and increase their fidelity in recognition.