Experiments with some algorithms that find central solutions for pattern classification

Abstract

In two-class pattern recognition, it is a standard technique to have an algorithm finding hyperplanes which separates the two classes in a linearly separable training set. The traditional methods find a hyperplane which separates all points in one class from all points in the other, but such a hyperplane is not necessarily centered in the empty space between the two classes. Since a central hyperplane does not favor one class or the other, it should have a lower error rate in classifying new points and is therefore better than a noncentral hyperplane. Six algorithms for finding central hyperplanes are tested on three data sets. Although frequently used in practice, the modified relaxation algorithm is very poor. Three algorithms which are defined in the paper are found to be quite good.