Eective supra-classi®ers for knowledge base construction

We explore the use of the supra-classi®er framework in the construction of a classi®er knowledge base. Previously, we introduced this framework within which labels produced by old classi®ers are used to improve the generalization performance of a new classi®er for a di€erent but related classi®cation task (Bollacker and Ghosh, 1998). We showed empirically that a simple Hamming nearest neighbor is superior to other techniques (e.g., multilayer perception (MLP), decision trees, Naive Bayes, Combiners) as a supra-classi®er. Here, we describe theoretically how the probability that the Hamming nearest neighbor supra-classi®er will predict the true target class approaches certainty at an exponential rate as more classi®ers are reused. The scalability of the Hamming nearest neighbor with large numbers of previously created classi®ers makes it a good choice as a supra-classi®er in the application of building a repository of domain knowledge organized as a classi®er knowledge base. Ó 1999 Elsevier Science B.V. All rights reserved.