A knowledge-oriented clustering technique based on rough sets

Abstract

Presents a knowledge-oriented clustering method based on rough set theory. The method evaluates the simplicity of classification knowledge during the clustering process and produces readable clusters reflecting the global features of objects. The method uses a newly-introduced measure, the indiscernibility degree, to evaluate the importance of equivalence relations that are related to the roughness of the classification knowledge. The indiscernibility degree is defined as the ratio of equivalence relations that gives a common classification to the two objects under consideration. The two objects can be classified into the same class if they have a high indiscernibility degree, even in the presence of equivalence relations which differentiate these objects. Ignorance of such equivalence relations is related to the generalization of knowledge, and it yields simple clusters that can be represented by simple knowledge. An experiment was performed on artificially created numerical data sets. The results showed that objects were classified into the expected clusters if modification was performed, whereas they were classified into many small categories without modification.