Seleção de protótipos com mapas-auto-organizáveis e entropia para sobreposição de classes e desbalanceamento de dados

• Ano de defesa: 2019
• Autor(a) principal: Rubbo, Márcio
• Orientador(a): Silva, Leandro Augusto da
• Banca de defesa: Menezes, Mario Olímpio de , Carvalho, André Carlos Ponce de Leon Ferreira de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Presbiteriana Mackenzie
• Programa de Pós-Graduação: Engenharia Elétrica
• Departamento: Escola de Engenharia Mackenzie (EE)
• País: Brasil
• Palavras-chave em Português: seleção de protótipos; mapas auto-organizáveis de Kohonen; K vizinhos mais próximos; sobreposição de classes; redução de dados; desbalanceamento; complexidade de dados
• Área do conhecimento CNPq: CNPQ::CIENCIAS EXATAS E DA TERRA
• Link de acesso: http://dspace.mackenzie.br/handle/10899/24499

The k nearest neighbor (kNN) is a traditional supervised classifier used in data mining tasks. However, when used in real applications, mainly in a dataset with class imbalance or class overlap, kNN suffers with problems in the task of data classification. In this work, we propose three prototype selection methods using self-organizing maps (SOM) and information entropy to increase the effectiveness of the kNN classifier in datasets with these conditions. Artificial datasets that simulate different conditions of data overlap and data imbalance were created and used together with public datasets to test the methods. Data complexity measures were used to identify data overlap and spatial distribution in the bases used in the experiment and a comparison was made with the results of the methods. The methods, named SOMEntropyHighFilter, SOMEntropyLowFilter and SOMEntropyHighLowFIlter, were able to increase the effectiveness of the kNN classifier in the artificial and real datasets used in the experiment, increasing the accuracy performance from datasets with imbalance or overlap problems.