MinOver Revisited for Incremental Support-Vector-Classification

Abstract

The well-known and very simple MinOver algorithm is reformulated for incremental support vector classification with and without kernels. A modified proof for its O(t 1/2) convergence is presented, with t as the number of training steps. Based on this modified proof it is shown that even a convergence of at least O(t 1) is given. This new convergence bound for MinOver is confirmed by computer experiments on artificial data sets. The computational effort per training step scales as O(N) with the number N of training patterns. © Springer-Verlag 2004.