D. P. Vetrov, D. A. Kropotov, N. O. Ptashko, “On the achievement of a compromise between the accuracy and stability of classifiers in the problem of choosing the best sound function for Bayesian learning”, Vestnik TVGU. Ser. Prikl. Matem. [Herald of Tver State University. Ser. Appl. Math.], 2009, Issue 15,Pages <nobr>47

PROBABILISTIC-STATISTICAL METHODS

On the achievement of a compromise between the accuracy and stability of classifiers in the problem of choosing the best sound function for Bayesian learning

D. P. Vetrov^a, D. A. Kropotov^b, N. O. Ptashko^a

^a Lomonosov Moscow State University, Faculty of Computational Mathematics and Cybernetics
^b Dorodnitsyn Computing Centre of the Russian Academy of Sciences, Moscow

Abstract: In the paper we show that RBF kernel selection in relevance vector machines (RVM) classifier requires extension of classifiers model. In new model integration over posterior probability becomes computationally unavailable. We propose a method of local evidence estimation which establishes a compromise between accuracy and stability of classifier.

Keywords: machine learning, bayesian framework, model selection, relevance vector machine.

UDC: 681.513.7

Received: 25.05.2009
Revised: 26.06.2009