Explorando abordagens de múltiplos rótulos por floresta de caminhos ótimos

• Ano de defesa: 2013
• Autor(a) principal: Pereira, Luís Augusto Martins [UNESP]
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Estadual Paulista (Unesp)
• Palavras-chave em Português: Computação - Matematica; Processamento de imagens - Tecnicas digitais; Reconhecimento de padrões; Floresta de caminhos ótimos; Arvores (Teoria dos grafos); Computer science Mathematics
• Link de acesso: http://hdl.handle.net/11449/122160

In conventional problems of pattern recognition, given a set of classes, each instance of the problem is associated with one and only one class. However, some real classification problems have instances that can be associated with more than one class at the same time, these problems are denoted as classification with multilabel. Among such problems, we highlight movies and music categorization, document classification, functional gene analysis etc. Nevertheless, the classification problems with multilabel are not directly treatable by conventional techniques, which explains the interest of pattern recognition community in these types of problems. Although many methods have been proposed in the literature, there is still much to be explored, especially in the use of novel conventional machine learning algorithms adapted or not to problems with multlabels. The Optimum-Path Forest (OPF) classifier is a supervised and deterministic algorithm applied to conventional classification problems, however, it has been not investigated in problems with multilabel. In this context, we investigated in this work the application of OPF-based classifiers on multilabel problems. We analyzed two versions of OPF-based classi ers: (i) the traditional one based on complete graph and (ii) the one based on k-nearest neighbors graph (OPFkNN). For manipulation of multilabel datasets, we used two transformation methods, the Binary Relevance and Label Powerset. We also proposed some changes in the training and classification phases of OPFkNN aiming to achieve better results when combined it with transformation methods. Experiments performed in seven public datasets showed that changes in OPFkNN improve outcomes. Comparison with the J48 classifier, ...