Donald Bren School of Information and Computer Sciences

Work in conceptual clustering has focused on creating classes from objects with a fixed set of features, such as color or size. In this paper we describe a system which uses relations between the objects being clustered as well as features of the objects to form a hierarchy tree of classes. Unlike previous conceptual clustering systems, this algorithm can define new attributes. Using relational information the system is able to find object classifications not possible with conventional conceptual clustering methods.