A Multi-Class Support Vector Machine Based on Scatter Criteria

Abstract

We re-visit Support Vector Machines (SVMs) and provide a novel interpretation thereof in terms of weighted class means and scatter theory. The gained theoretical insight can be translated into a highly efficient extension to multi-class SVMs: mScatter-SVMs. Numerical simulations reveal that more than an order of magnitude speed-up can be gained while the classification performance remains largely unchanged at the level of the classical one vs. rest and one vs. one implementation of multi-class SVMs.