Adjusted estimation for the combination of classifiers

Article type: Research Article

Authors: Mertens, Bart J.A.^a | Hand, David J.^b

Affiliations: [a] Department of Biostatistical Sciences, Academic Medical Centre, University of Amsterdam, 1100 DD Amsterdam, The Netherlands. E-mail: b.j.mertens@amc.uva.nl | [b] Department of Statistical Sciences, Imperial College, London SW7 2BZ, UK. E-mail: d.j.hand@ic.ac.uk

Abstract: An algorithm is proposed which adaptively and simultaneously estimates and combines classifiers originating from distinct classification frameworks for improved prediction. The methodology is developed and evaluated on simulations and real data. Analogies and similarities with generalized additive modeling, neural estimation and boosting are discussed. We contrast the approach with existing Bayesian model averaging methods. Areas for further research and development are indicated.

Keywords: combining information, model uncertainty, pattern recognition, classification and regression, automated data analysis

DOI: 10.3233/IDA-2000-4206

Journal: Intelligent Data Analysis, vol. 4, no. 2, pp. 165-179, 2000