Geração de conjuntos consistentes de regras para classificação multirrótulo com algoritmo evolutivo multiobjetivo

• Ano de defesa: 2020
• Autor(a) principal: Miranda, Thiago Zafalon
• Orientador(a): Cerri, Ricardo
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC
• Palavras-chave em Português: Classificação multirrótulo; Interpretabilidade; Regras consistentes; Otimização multiobjetivo; Algoritmos evolutivos
• Palavras-chave em Inglês: Multi-label classification; Interpretability; Consistent rules; Multi-objective optimization; Evolutionary algorithms
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO::SISTEMAS DE INFORMACAO
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/13067

Multi-label classification a machine learning task whose objective is to generate models capable of learning relationships between descriptive characteristics of objects and the sets of classes to which such objects belong. In certain applications, it is important for the models to be interpretable so that their users can trust it or so that its predictions can be explained. In this research, we investigated the generation of multi-label classification models based on consistent sets of rules. We proposed an evolutionary algorithm and two auxiliary algorithms that guide the rule generation process, ensuring that the rules created were consistent with each other. A set of rules is consistent if whenever multiple rules covers an object, such rules predict the same set of classes. The proposed evolutionary algorithm utilized multi-objective optimization techniques to generate collections of classification models that offer different compromises between interpretability and predictive power. Experiments were conducted with the proposed algorithms and with algorithms from the literature and, based on statistical analysis, we concluded that the generated models were, in terms of interpretability, superior to those generated by literature's algorithms and, in terms of predictive power, they were comparable to most.