A new Kernel to use with discretized Temporal Series

Abstract

In this papera new kernel, from statistical learning theory, is proposed to work with symbols chains (words) obtained from a discretization procedure of a continuous feature. Meanwhile the exact defini­tion of the discretization is not strictly necessary, it must ever exist either, a distance or a similarity measure between symbols in a certain alphabet (a set of symbols). The proposed kernel is a generalization of a dot product in a vector space, not necessarily provided of any mathematical structure, that will allows to establish a similarity measure between objects of the alphabet. This kernel is applied on a set of television shares obtained from the seven main television stations in Andalusia (Spain). A comparative study for clas­sification purposes is done, and the associated pa­rameter selection is studied.