Uso e estabilidade de seletores de variáveis baseados nos pesos de conexão de redes neurais artificiais

• Ano de defesa: 2021
• Autor(a) principal: Costa, Nattane Luíza da
• Orientador(a): Barbosa, Rommel Melgaço
• Banca de defesa: Barbosa, Rommel Melgaço, Lima, Márcio Dias de, Lins, Isis Didier, Costa, Ronaldo Martins da, Leitão Júnior, Plínio de Sá
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Goiás
• Programa de Pós-Graduação: Programa de Pós-graduação em Ciência da Computação (INF)
• Departamento: Instituto de Informática - INF (RG)
• País: Brasil
• Palavras-chave em Português: Redes neurais artificiais; Seleção de variáveis; Pesos de conexão; Estabilidade; Ranking de importância
• Palavras-chave em Inglês: Artificial neural network; Feature selection; Connection weights; Stability
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::SISTEMAS DE COMPUTACAO
• Link de acesso: http://repositorio.bc.ufg.br/tede/handle/tede/11396

Artificial Neural Networks (ANN) are machine learning models used to solve problems in several research fields. Although, ANNs are often considered “black boxes”, which means that these models cannot be interpreted, as they do not provide explanatory information. Connection Weight Based Feature Selectors (WBFS) have been proposed to extract knowledge from ANNs. Most of studies that have been using these algorithms are based on just one ANN model. However, there are variations in the ANN connection weight values due to the initialization and training, and consequently, leading to variations in the importance ranking generated by a WBFS. In this context, this thesis presents a study about the WBFS. First, a new voting approach is proposed to assess the stability of the WBFS, i.e, the variation in the result of the WBFS. Then, we evaluated the stability of the algorithms based on multilayer perceptron (MLP) and extreme learning machines (ELM). Furthermore, an improvement is proposed in the algorithms of Garson, Olden, and Yoon, combining them with the feature selector ReliefF. The new algorithms are called FSGR, FSOR, and FSYR. The experiments were performed based on 28 MLP architectures, 16 ELM architectures, and 16 data sets from the UCI Machine Learning Repository. The results show that there is a significant difference in WBFS stability depending on the training parameters of the ANNs and depending on the WBFS used. In addition, the proposed algorithms proved to be more effective than the classic algorithms. As far as we know, this study was the first attempt to measure the stability of WBFS, to investigate the effects of different ANN training parameters on the stability of WBFS, and the first to propose a combination of WBFS with another feature selector. Besides, the results provide information about the benefits and limitations of WBFS and represent a starting point for improving the stability of these algorithms.