Classification on dissimilarity data : a first look

In a dissimilarity (distance) data each pair of objects is characterized by a value which expresses the magnitude of difference between them. This type of data can be now classified using various approaches, provided that a new object is represented by its distances to the training samples. This paper discusses a number of possibilities to tackle such a classification problem. Two types of methods are investigated: the feature-based (i.e. interpreting the distance data as a feature space) and rank-based decision rules. Experiments conducted on real datasets demonstrate that the feature-based classifiers often outperform the rank-based ones. The normal-based decision rules perform well, since summation-based distances (frequently appearing in practice) are, under general conditions, approximately normally distributed. In addition, also the support vector classifier achieves a high accuracy, particularly in distance spaces of a very high dimensionality.