A Compact Representation of Multiscale Dissimilarity Data by Prototype Selection

Abstract

Multiscale information provides an opportunity to improve the outcomes of data analysis processes. However, if the multiscale information is not properly summarized in a compact representation, this may lead to problems related to high dimensional data. In addition, in some situations, it is convenient to define dissimilarities directly for the multiscale data obtaining in this way a multiscale dissimilarity representation. When these dissimilarities are specifically designed for the problem, it is even possible that they do not fulfill metric requirements. Therefore, standard statistical analysis techniques may not be easily applicable. We propose a new method to combine non-metric multiscale dissimilarities in a compact representation which is used for classification. The method is based on the extended multiscale dissimilarity space and prototype selection, which allows us to handle the potentially non-metric nature of the dissimilarities and exploit the multiscale information at the same time. This is achieved in such a way that the most informative examples per scale are selected. Experimental results show that the approach is promising since it finds a better trade-off in accuracy and efficiency than its counterpart approaches.