The Role of Prototypicality in Exemplar-Based Learning

Abstract

This paper examines the role of prototypicality in exemplarbased concept learning methods. It proposes two approaches to prototypicality: a shared-properties approach, and a similarity-based approach, and suggests measures that implement the different approaches. The proposed measures are tested in a set of experiments. The results of the experiments show that prototypicality serves as a good storing filter in storage reduction algorithms; combining it in algorithms that store all the training set does not improve significantly the accuracy of the algorithm. Finally, prototypicality is a useful notion only in a subset of the domains; a preliminary examination of those domains and their characteristics is proposed.