Estudos em estimação de densidade por Kernel: métodos de seleção de características e estimação do parâmetro suavizador

Abstract

Function induction problems are frequently represented by affinity measures between the elements of the inductive sample set, being kernel matrices a well known one. This work have as objective obtain information of the relations between data from the calculated kernel matrix, starting from the hypothesis that those geometric relations are coherent with known labels. Univariate and multivariate feature selection methods that use kernel density estimation (KDE) were proposed. Methods for perform estimation of kernel width, based at the geometric coherence between label and problem geometry, were also proposed. To assess the relation of data structure with the labels, a classifier based on kernel density estimation (KDE) was used and the performance of the proposed methods was compared with others known from literature. To the databases tested, the performance of the proposed methods were similar to the ones in the literature. Results indicates that is practicable selecting models through the direct calculation of densities and the geometry from the class separation.