Agregação de classificadores neurais via integral de Choquet com respeito a uma medida fuzzy

• Ano de defesa: 2016
• Autor(a) principal: Pacheco, André Georghton Cardoso
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Espírito Santo; BR; Mestrado em Informática; Centro Tecnológico; UFES; Programa de Pós-Graduação em Informática
• Palavras-chave em Português: Data classification; Ensemble of classifiers; Choquet integral; Fuzzy measure; Deep learning; Choquet, Integrais de; Aprendizagem profunda (Computação); Classificação de dados; Elencos de classificadores; Medida fuzzy; Aprendizado profundo; Algoritmos difusos; Redes neurais (Computação); Análise de componentes principais; Ciência da Computação; 004
• Link de acesso: http://repositorio.ufes.br/handle/10/6811

Data classification appears in many real-world problems, e.g., recognition of image patterns, differentiation among species of plants, classifying between benign and malignant tumors, among others. Many of these problems present data patterns, which are difficult to be identified, thus requiring more advanced techniques to be solved. Over the last few years, various classification algorithms have been developed to address these problems, but there is no classifier able to be the best choice in all situations. So, the concept of ensemble systems arise, which more than one methodology is used together to solve a particular problem. As a simple and effective methodology, ensemble of classifiers have been applied in several classification problems, aiming to improve performance and increase reliability of the final result. However, in order to improve the classification accuracy, an affective aggregation of classifiers must be performed. In this work, we present two contributions: first, we describe three classifiers based on neural networks, a multilayer feedforward trained by Levenberg-Marquardt algorithm; an extreme learning machine (ELM); and a discriminative restricted Boltmann machine (DRBM). Furthermore, we use conventional classifier k-nearest neighbors (KNN). Next, we propose an aggregation methodology to ensemble of classifiers using Choquet integral with respect to a fuzzy measure obtained by principal component analysis (PCA). Then, we apply this methodology to aggregate the classifiers performed to conventional benchmarks, for large database and the results are promising.