Taxonomy of classifiers based on dissimilarity features

Abstract

A great number of linear and nonlinear classification algorithms can follow from a general representation of discriminant function written as a weighted sum of kernel functions of dissimilarity features. Unified look at the algorithms allows obtaining intermediate classifiers which after tuning of the weights and other model's parameters can outperform popular dissimilarity based methods. Simulations with artificial and real world data sets revealed efficiency of single layer perceptron trained in a special way. © Springer-Verlag Berlin Heidelberg 2005.