Dissimilarity based metric for data classification based on Support Vector Classifiers

A class is a concept of a set of objects possessing similar characteristics. This implies that the notion of
“similarity/dissimilarity” is as fundamental as of “feature”, since it is the similarity which groups the objects together in a class. The distance representation is most commonly used as a dissimilarity measure because is usually the simplest to calculate. Dissimilarity-based pattern recognition offers new possibilities for building classifiers on a distance representation such as kernel methods or the k nearest neighbors (kNN) rule. The goal of this work is to expand and ameliorate the advantageous and rapid adaptive approach to learn only from dissimilarity representations developed by Guerin and Celeux [GUE 01] by using the effectiveness of the Support Vector Machines algorithm for real-world classification tasks. This method can be an alternative approach to the known methods based on dissimilarity representations
such as Pekalska’s dissimilarity classifier [PEK 05], Haasdonk’s kernel-based SVM classifier [HAA 04] and to kNN classifier and can be as effective as them in terms of accuracy for classification. Practical examples on artificial and real data show interesting behavior compared to other dissimilarity-based methods.

• Agata Manolova
• Anne Guerin-Dugue

Rencontres de la Société Francophone de Classification (SFC), 2009

http://www.sfc-classification.net/spip.php?article5