An Empirical Comparison of Kernel-Based and Dissimilarity-Based Feature Spaces

The aim of this paper is to find an answer to the question: What is the difference between dissimilarity-based classifications(DBCs) and other kernel-based classifications(KBCs)? In DBCs [11], classifiers are defined among classes; they are not based on the feature measurements of individual objects, but rather on a suitable dissimilarity measure among them. In KBCs [15], on the other hand, classifiers are designed in a high-dimensional feature space transformed from the original input feature space through kernels, such as a Mercer kernel. Thus, the difference that exists between the two approaches can be summarized as follows: The

distance

kernel of DBCs represents the discriminative information in a relative manner, i.e. through pairwise dissimilarity relations between two objects, while the

mapping

kernel of KBCs represents the discriminative information uniformly in a fixed way for all objects. In this paper, we report on an empirical evaluation of some classifiers built in the two different representation spaces: the dissimilarity space and the kernel space. Our experimental results, obtained with well-known benchmark databases, demonstrate that when the kernel parameters have not been appropriately chosen, DBCs always achieve better results than KBCs in terms of classification accuracies.