An Entropic Approach to Dimensionality Reduction in the Representation Space on Discrete Processes

This paper generalizes a method for class discrimination with the purpose of formulating dynamic random models starting from sample information.Such technique pursues,through the study of the eigenvalues,the reduction of the dimensionality in the representation space.We also describe an algorithm that allows the reconstruction and the approximation by dimensionality reduction for the original information.An illustration with a theoretical model reveals the great compression power produced by this scheme,as well as the goodness of the approximations by dimensionality reduction.Two class case is briefly discussed and an algorithm for reconstruction and classification is suggested.An application to meteorological data shows results similar to the one class case as far as dimensionality reduction goes.A reasonable classification rate is also obtained.