Nature-Inspiration on Kernel Machines: Data Mining for Continuous and Discrete Variables

Abstract

Kernel Machines, like Support Vector Machines, have been frequently used, with considerable success, in situations in which the input variables are given by real values. Furthermore, the nature of this machine learning algorithm allows extending its applications to deal with other kinds of systems with no vectorial information such as facial images, hand written texts, micro-array gene expressions, or protein chains. The behavior of a number of systems could be better explained if artificial infinite-precision variables were replaced by qualitative variables. Hence, the use of ordinal or interval scales on input variables would allow kernels to be defined for nature-inspired systems directly. In this contribution, two new kernels are designed for applying kernel machines to such systems described by qualitative variables (orders of magnitude or intervals). In addition, the structure of the feature space induced by this kernel is also analyzed.