Modelo alfa normal assimétrico multivariado para redes de classificação

Abstract

In this Thesis we expose the proposition of a new class of probability distributions, the so called alpha skew normal multivariate, an extension of the univariate Normal Alpha distribution, introduced by Elal-Olivero (2010). It can accommodates up to two modes and generalizes the distribution proposed by Elal-Olivero in its marginal components. In addition, we apply this new distribution in the construction of two new data mining methods for classi cation. The procedures developed here increment the predictive ability of the classi cation in the presence of asymmetric and / or bimodal data. The results indicate that the new proposal is signi cantly more appropriate than the usual modeling by classical normal distribution, and is also suitable for datasets without the presence of asymmetry. In this thesis it is shown, using real and synthetic data, the procedures of construction, estimation and validation for the new probability distribution and for probabilistic networks for binary classi cations, particularly for the k-dependence probabilistic networks.