Conceptual Clustering in Structured Domains: A Theory Guided Approach

Abstract

In the paper, ULS, a conjunctive clustering system working on structured descriptions, is described. It uses background knowledge defined as a theory about the classification goals. It involves learning category structure from previous information concerning the features and the domain in order to simplify, at least from a computational point of view, the clustering process. ULS explicitly couples clustering and characterization: the task of clustering is performed adopting a similarity based approach while characterization aims at intensionally defining concepts. The background knowledge is adopted in clustering, in order to partition the initial space by the definition of class prototypes, and in characterization, in order to define some heuristics useful to guide the search process in concept generalization.