A ordenação das variáveis no processo de otimização de classificadores bayesianos: uma abordagem evolutiva

• Ano de defesa: 2007
• Autor(a) principal: Santos, Edimilson Batista dos
• Orientador(a): Hruschka Júnior, Estevam Rafael
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC
• País: BR
• Palavras-chave em Português: Redes Bayesianas; Aprendizado do computador; Algoritmos genéticos
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/361

Classification is a basic task in data analysis and pattern recognition that requires the construction of a classifier. The induction of classifiers from data sets is an important problem in machine learning. Numerous approaches to this problem are based on various representations such as decision trees, neural networks, decision graphs, and rules. However the interest in Bayesianos methods for classification has grown sufficiently. Bayesian Networks (BNs) learning algorithms can be used to induce Bayesian classifiers. However, BNs learning from data is known to be a NP problem and does not have computational methods capable to identify to the best solution for all the application problems. A very common restriction when learning a BN is the definition of a previous Variables Ordering (OV). The OV represent the possible relationships between the variables in the formation of the structure of BN that describes the problem. Using an adequate OV, learning algorithms are capable to find a solution more efficient. Therefore, this work proposes hybrid approaches to help the process of learning a BN from data for classification. The proposed methods named VOGA, VOGAC e VOEA uses Evolutionary Algorithms to optimize the BN learning process by means of the identification of an adequate variables ordering. These methods use information about the class variable when defining the most suitable variable ordering. Experiments performed in a number of datasets revealed that methods are promising