Prototype Selection for Classification in Standard and Generalized Dissimilarity Spaces

A common way to represent patterns for recognition systems is by feature vectors lying in some space. If this representation is based only on the predefined object features, it is independent of the other objects. In contrast, a dissimilarity representation of objects takes into account the relations between them by some measure of resemblance (e.g. dissimilarity). The nearest neighbour (1-NN) is a dissimilarity-based classifier that has shown to be very competitive for several pattern recognition problems. Classification results on dissimilarity spaces spanned by dissimilarities to prototypes can reach or improve the 1-NN results in terms of accuracy and computational efficiency. This is possible if a small set of prototypes is selected with similar discriminative power than the complete set of initial prototypes. How to obtain this set has been studied by researchers in the area of dissimilarity representations and graph representations by means of prototype selection methods. In this chapter we present an overview and a discussion of different approaches proposed in the literature on this topic.