Análise discriminante via distribuições preditivas aproximadas por estimadores por função núcleo

Abstract

Pattern Recognition and Classification problems are importantes in a variety of science fields, such as biology, psychology, medice, computer vision and etc. However, the problem is not so easy to solve when the true probability distribution of data is unknown. In this work, we combine the Kernel density estimation method with a Bayesian approach and propose a new method for classification problems using Discriminant Analisys via Approximate Predictive Distribution. Simulation studies and application in data sets widely used in literature, were conducted as an assessment of the proposed methods. The results showed that the performace of the proposed methods are competitive, and in some cases significantly better, with classical methods of literature, Linear Discriminant Analisys (LDA), Quadratic Discriminant Analisys (QDA) and Naive Bayes Discriminant Analisys with Normal distribution (NNBDA).