Uma metodologia de caracterização de serviços de mineração de dados

• Ano de defesa: 2005
• Autor(a) principal: Leonardo Chaves Dutra da Rocha
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; UFMG
• Palavras-chave em Português: Mineração de Dados; Serviços; World Wide Web (Sistema de recuperação da informação); Mineração de dados (Computação); Sistemas de recuperação da informação; Tecnologia da informação; Data mining (Sistemas de recuperação da informação)
• Link de acesso: http://hdl.handle.net/1843/SLBS-6GUL2M

Web services are becoming a standard for deploying a large spectrum of applications using the Internet. One example of such application is data mining, which aims to extract useful information from large sets of data. A successful data mining service must fulfill both interaction requirements of a data mining task, and the intensive processing and large storage requirements usually associated with the techniques employed. In this paper we present a methodology for characterizing computationally intensive Web services, in particular data mining services. Our characterization methodology focuses on both the interactive and non-interactive sides of the service, as well as their relationships. We applied our methodology on an actual data mining service, Tamanduá. The results show that there is a high variability among users in terms of their behavior, but they act similarly with respect to the nature of the data mining tasks they request and how they analyze the results. Our results also show that we are able to divide the users into two distinct groups, one that uses the system in a sequential fashion and other that presents an asynchronous behavior, placing a bigger demand on the system. These results not only show the applicability of our approach, but also open new directions in terms of system support mechanisms for such Web services. Further, we are also able to determine statistical distributions that may be used for generating synthetic workloads that would allow a detailed performance analysis of the system.