Learning Predictive Qualitative Models with Padé
Jure Žabkar, Martin Možina, Ivan Bratko, Janez Demšar · Repository of the University of Ljubljana (University of Ljubljana) · 2011
provide insight into how a change of a certain input variable affects the output within a context of other inputs. Although people usually reason qualitatively, machine learning has mostly ignored this type of model. We present a new approach to learning qualitative models from numerical data. We describe Pade, a suite of methods for estimating partial derivatives of unknown sampled target functions. We show how to build qualitative models using standard machine learning algorithms by replacing the output variable with signs of computed derivatives. Experiments show that the developed methods are quite accurate, scalable to high number of dimensions and robust with regard to noise. Povzetek: Predstavljena je nova metoda za uy cenje iz kvalitativnih podatkov, imenovana Pade. Temelji na ocenjevanju parcialnih odvodov neznane vzory ciljne funkcije.