Um método determinístico para redução de conjuntos de dados genéricos grandes utilizando Difusão Geométrica Markoviana

Abstract

Instance reduction techniques can be used to enable the exploration of large sets of general data to obtain information. Extended Markovian Geometric Diffusion (DGM-E) is an instance selection method for large generic datasets, which do not have an intrinsically defined geometric structure. For this reason, the DGM-E has a data structuring algorithm, which is responsible for scaling the data, partitioning the data space and determining the neighborhood of each element, in a probabilistic way, among the elements of the same partition. This work presents a new deterministic approach to neighborhood construction, which infers the k-nearest neighbors of each element. The DGM-E instance reduction algorithm is able to infer the degree of representativeness of each instance of the dataset. From this data, the element reduction technique can be applied, with which less important instances are removed.