Concept Formation through Clustering

Abstract

Clustering can be viewed as a basic approach to concept learning. Previous methods of conceptual clustering select, based on the criteria of maximizing the quality of clustering, appropriate attributes from the a priori given set of ones in order to produce a generalized description for each concept. We here propose a new learning model CNC, consisting of numerical and conceptual parts, which can form concepts by the cooperative tasks of two parts even under the circumstance of giving no prespecified attributes. The numerical part in CNC produces similarity-based clusters by the quantitative treatment of the objects and assists the conceptual part in getting the generalized descriptions for specifying clusters and discriminating different clusters, conceptually. Needed attributes for the descriptions are generated in the course of learning by forming recursively a conceptual structure for simplified versions of the original objects. Computational experiments are also examined to see how CNC works well for the real world color data. As a result several concepts, to which we give the names, red, pink, orange, brown, etc., were formed together with a few color attributes for characterizing them.