Merging multiple sources of information using dissimilarity based approaches

Abstract. The ways distances are computed (the metric used) or mea-sured (by mean of different sources) enable us to have different rep-resentations of the same objects. Therefore we need either to make a selection or to combine them. A combination of differently measured (or computed) dissimilarities can occur at different stages of a pattern recognition system, e.g. using the outputs of classifiers built on each of them separately but also by combining the various dissimilarity directly. The key point of classifier combination lies either in a proper averag-ing over different experts/sources or in an integration of different and hopefully complementary approaches. In this paper we want to focus on possible ways of merging different sources of informations given by dif-ferent dissimilarity representations. The combination step is employed in a selection or weighted average of the data used for our classifica-tion task. A simple averaging of these matrices is shown ([1], [2], [3]) to allow for classification accuracies that outperform the single ones. We compare this approach here with a dissimilarity forward selection and other techniques based on the learning of weights of linear and quadratic forms. The more advanced forms of combination of such multiple repre-sentations can lead to a better representation of the underlying data and therefore better classification accuracies but this does not hold always. Our general conclusion is that combining given distance matrices before any training stage is apparently always worthwhile.