BR: A new method for computing all typical testors

Abstract

Typical testors are very useful in Pattern Recognition, especially for Feature Selection problems. The complexity of computing all typical testors of a training matrix has an exponential growth with respect to the number of features. Several methods that speed up the calculation of the set of all typical testors have been developed, but nowadays, there are still problems where this set is impossible to find. With this aim, a new external scale algorithm BR is proposed. The experimental results demonstrate that this method clearly outperforms the two best algorithms reported in the literature. © 2009 Springer-Verlag Berlin Heidelberg.