Discovering and using knowledge from unsupervised data

Though most knowledge discovery methods have been developed for supervised data, the task of finding knowledge from unsupervised data often arises in real-world situations. Without feedback about the appropriateness of discovered knowledge as in supervised systems, techniques for unsupervised knowledge discovery are essentially different and still much less developed than those for supervised discovery. In this paper we present a method for discovering and using classificatory knowledge from unsupervised data. We first extended the classical view on concepts commonly used in the framework of the Galois lattice by combining it with the prototype and exemplar views, and develop an algorithm for inducing concept hierarchies. We then introduce a procedure that combines matching approaches in inductive learning with case-based reasoning in order to classify unknown cases, using discovered knowledge. We present the implementation of the method as an interactive system. An experimental comparative study of some knowledge discovery systems, in terms of knowledge description and prediction, shows advantages and application potential of the method in decision-making. (C) 1997 Elsevier Science B.V.