Algoritmos integrados para classificação de dados com atributos categóricos

• Ano de defesa: 2019
• Autor(a) principal: Leite, Gabriel Matos Cardoso
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Rio de Janeiro; Brasil; Instituto Alberto Luiz Coimbra de Pós-Graduação e Pesquisa de Engenharia; Programa de Pós-Graduação em Engenharia de Sistemas e Computação; UFRJ
• Palavras-chave em Português: Classificação supervisionada; Neighbourhood component analysis; Dados categóricos; CNPQ::ENGENHARIAS
• Link de acesso: http://hdl.handle.net/11422/14059

Pattern classification on categorical and mixed data is a challenge to be surpassed. The increase in the amount of data being generated demands classifiers able to deal with different types of data. This work proposes algorithms for supervised classification on categorical and mixed data. Such algorithms are elaborated from integration between classifiers and ways of coding categorical features into continuous features. Mixed data is a set of observations with categorical features along with continuous features. Treating observations with categorical features properly allows the use of a huge number of databases containing categorical features. The approach proposed in order to handle categorical features and permit classification methods to be applied on such data, is a result of integration in pairs between the encodings Target Encoding (TE), One-hot, Naive and classifiers Neighbourhood Componente Analysis (NCA), Support Vector Machine (SVM), k-Nearest Neighbors (kNN). The behavior of the encodings chosen, and the performance of the presented algorithms are analyzed on synthetic databases and real databases, respectively. In order to evaluate the performance of the presented algorithms, an analysis was made on all results obtained. This analysis was made using crossvalidation techniques, k-fold and a test set with unseen observations. Moreover, inferential statistics techniques were used to identify evidences of differences among integrated algorithm’s accuracies on each dataset. The experimental planning proposed indicated that the integration built by NCA classifier and TE encoding (NCA+TE) turned up to be more competitive when compared to the other algorithms.