Exploração do uso de short-term memory na construção de métodos de acesso métricos dinâmicos sobre a perspectiva de diferentes políticas de divisão de nós

Abstract

This work proposes the development of a new approach using a structure called shortterm memory for the construction of dynamic MAMs derived from Slim-Tree. The goal is to minimize the overlap of nodes in the structure and, consequently, to optimize similarity queries. The strategy is based on postponing the indexing process to allow inserting, in the same index entry, new objects with closer distances. In addition, different node split policies were employed, which allowed to measure the impacts of these in what refers to the quality of the resulting structures, mainly with respect to the degree of overlap. The proposed new methods were evaluated by real and synthetic datasets and compared with the original MAM Slim-tree. The results of the experiments were significant with respect to the quality of the resulting trees, which consequently impacted on better efficiency in similarity query operations, significantly reducing the number of distance calculations, the number of disk page accesses and the execution time of k-nearest neighbors.