Finding dependencies in data based on methods of satisfying table con-straints

Abstract

The work deals with the search for a special type of regularities in data, called frequent patterns. A frequent pattern is understood as a certain set of attributes that characterizes a sufficiently large number of objects of the training sample. There are many methods for pattern discovery, but they usually do not allow flexible consideration of necessary requirements for their type. Taking into account the new conditions that the desired patterns must meet leads in practice to a time-consuming modification of used algorithms and a decrease in computing performance. This article proposes a new approach based on the constraint programming paradigm, which is free from the listed disadvantages. The approach is based on the original way of presenting the training sample using specialized table constraints – compressed D-type tables, on the author's method of backtracking, as well as on specialized reduction rules for table constraints. Particular attention is paid to solving the closed patterns discovery problem, which is included as part of the solution of all machine learning problems considered in the work, which means taking into account additional requirements for the type of patterns. As additional requirements to the type of pattern, constraints on the frequency of occurrence of a closed pattern, as well as conditions for the occurrence of some element (attribute) into the pattern, are considered. To the basic rules for the reduction of compressed D-type tables, rules are added that take into account the interesting attributes of the analyzed patterns. The advantage of the approach is that the taking into account and analyzing new constraints makes it possible to speed up the calculation process.