Focused Concept Formation

Abstract

This chapter reviews the incremental concept formation systems COBWEB (Fisher, 1987) and CLASSIT (Gennari, Langley L. Fisher), and also CLASSIT-2. For COBWEB and CLASSIT, recognition of an instance occurred when all available attributes were used to classify the instance into some category. A better approach would be to recognize an instance based on only a small number of attributes. Additionally, a clustering system should be able to focus attention on some subset of attributes that are most salient for a given classification problem. These attributes should be inspected in sequence. The incremental algorithm used by both COBWEB and CLASSIT is only a partial specification of the clustering method. CLASSIT-2 extends the framework of incremental concept formation to include a mechanism for attention. This extension is well-integrated with the existing framework. The ability to focus attention on selected attributes of an instance is both an important step toward a model of human concept formation, and a useful method for improving efficiency without losing predictive ability.