Abstract
We reformulate the support vector machine approach to classification and regression problems using a different methodology than the classical ‘largest margin’ paradigm. From this, we are able to derive extremely simple quadratic programming problems that allow for general symbolic solutions to the classical problems of geometric classification and regression. We obtain a new class of learning machines that are also robust to the presence of small perturbations and/or corrupted or missing data in the training sets (provided that information about the amplitude of the perturbations is known approximately). A high performance framework for very large-scale classification and regression problems based on a Voronoi tessellation of the input space is also introduced in this work. Our approach has been tested on seven benchmark databases with noticeable gain in computational time in comparison with standard decomposition techniques such as SVM light .
Acknowledgements
The authors wish to acknowledge NSF under grant DMI0427966 for providing financial support for this work. The opinions expressed herein are those of the authors and not necessarily those of the NSF.
Notes
E-mail: [email protected]. URL: http://www.robingilbert.com
URL: http://www.lois.ou.edu
A kernel is a continuous function k:E×E→R on any set E that is symmetric and positive semi-definite.
A RKHS
is a Hilbert space such that for every
x
∈E the linear map f∈
↦f(
x
)∈
is continuous. Given a kernel k, it is possible to construct a unique RKHS with k as the reproducing kernel.
The expected value of the average number of observations misclassified by
.