t-SNE paralelo: uma técnica paralela para redução de dimensionalidade de dados aplicada em Cidades Inteligentes

• Ano de defesa: 2020
• Autor(a) principal: Lopes, Maximiliano Araújo da Silva
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Rio Grande do Norte; Brasil; UFRN; PROGRAMA DE PÓS-GRADUAÇÃO EM ENGENHARIA ELÉTRICA E DE COMPUTAÇÃO
• Palavras-chave em Português: Big Data; Cidades Inteligentes; Programação paralela; Redução de dimensionalide; t-SNE
• Link de acesso: https://repositorio.ufrn.br/handle/123456789/31522

The urban population has been growing rapidly around the world. To adapt to this population increase, mayors need to change the way they manage these large cities. Thus, the concept of smart cities gains strength and comes to change the way of life of the world population. The investment in this concept of smart cities is aimed at improving people’s management and quality of life. The biggest challenges in these systems are linked to the processing, visualization and analysis of the generated data, since when they work connected these systems generate a large mass of data, called Big Data, which need to be treated differently from conventional systems. For the visualization of the data, a device that can be used are the techniques for reducing the dimensionality, which bring the data from one n dimension to two or three dimensions, thus being perceptible to human eyes. One of the main problems involving Dimensionality Reduction techniques is related to the processing time, which makes them practically unfeasible to be applied to large masses of data. In this thesis, a way to decrease the processing time of these algorithms is presented, by parallelizing the t-SNE algorithm. An analysis was performed on each part of the algorithm, verifying which sections could be parallelized and which sections would continue with their conventional processing. In this way, the parallelized algorithm showed better results than its conventional version, presenting itself as a more efficient and effective technique in Reducing the Dimensionality of data in order to optimize their visualization and analysis.