Generalizing dissimilarity representations using feature lines

Abstract

A crucial issue in dissimilarity-based classification is the choice of the representation set. In the small sample case, classifiers car pable of a good generalization and the injection or addition of extra information allow to overcome the representational limitations. In this paper, we present a new approach for enriching dissimilarity representations. It is based on the concept of feature lines and consists in deriving a generalized version of the original dissimilarity representation by using feature lines as prototypes. We use a linear normal density-based classifier and the nearest neighbor rule, as well as two different methods for selecting prototypes: random choice and a length-based selection of the feature lines. An important observation is that just a few long feature lines are needed to obtain a significant improvement in performance over the other representation sets and classifiers. In general, the experiments show that this alternative representation is especially profitable for some correlated datasets. © Springer-Verlag Berlin Heidelberg 2007.