Nacluster: resolvendo entidades em larga escala a partir de múltiplos catálogos de astronomia

• Ano de defesa: 2016
• Autor(a) principal: Freire, Vinícius Pires de Moura
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Resolução de entidades; Casamento de catálogos; Big data
• Link de acesso: http://www.repositorio.ufc.br/handle/riufc/39385

Astronomy surveys use powerful instruments to browse the sky and identify objects of interest within the surveyed region. Sky objects are individually characterized with spatial coordinates, identifying their position in the sky, in addition to other descriptive attributes. Composing an integrated view of the sky based on catalogues produced by different surveys faces a hard problem of matching objects that have been captured by different telescopes. Due to variations on capturing instruments calibration, the sky position of a single sky object may vary from a catalog to the other. Moreover, in particular dense regions of the sky this problem is exacerbated by a huge number of candidate matches for each given object. Traditional approaches for dealing with this problem use a threshold distance of ε to reduce the number of matching candidates. Additionally, they adopt a pairwise approach for matching n catalogues inferring transitivity among matches, which not always hold. In this thesis, we present NACluster a non-supervised clustering algorithm for dealing with sky object matching in multiple catalogues. NACluster matching strategy extends the traditional k-means clustering algorithm by relaxing the number k of cluster (i.e. matched sky objects). We propose the ParallelNACluster, a parallel version of NACluster that takes advantage of partitioning the input data, and accept large volumes of data using a set of conventional hardware. In addition, we propose the SCIBoundary, a new strategy for matching neighboring stars placed in different data partitions. The strategy leads to equivalent solutions in both NACluster and ParallelNACluster. In this thesis, we present also the AODP, a workflow to perform partitioning of data on local disk and run their matching in distributed environment via ParallelNACluster. Our experiments show the efficiency of centralized and parallel strategies, as well as the efficiency of SCIBoundary in the treatment of borders.