Extração de forma compacta de regras Fuzzy de uma rede bayesiana

• Autor(a) principal: Hsien, Yin
• Data de Publicação: 2010
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Institucional da UFSCAR
• Texto Completo: https://repositorio.ufscar.br/handle/ufscar/453
• topic: Data mining (Mineração de dados) - métodos estatísticos; Redes Bayesianas; Lógica nebulosa; Seleção de atributos; Inteligência artificial; CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO

The decision support tools are importants in many domains of our society. But there is a need to make users understand the decisions made by those tools in order to increase the faith of the users on the results. In the literature, Bayesian networks are considered as a probabilistic classification system with good performance. But it still need a better presentation of it results to make it more understandable to the users. In other hand, the fuzzy logic ofer potential to deal with imprecision and uncertainty, as well as a linguistic representation, which facilitates user's understanding. The combination of Bayesian networks and fuzzy logic is proposed by the method BayesFuzzy, which make use of fuzzy rules as a form of explanation of a Bayesian network, it aims to obtain a decision support tool with good performance and easy to be understood by users. So we are proposiing the method Pruned BayesFuzzy (PBF), it is a BayesFuzzy incorporated with minimum certainty degree, default rule and Rule Post-Pruning as a form to select the most important rules for classification between all the rules generated, it also simplifies those rules. The results of PBF show an improvement in understanding but a loss in correct classification rate. But the improvement in understanding is promising enough to further research and enhance of the PBF. Then beside PBF, we also propose the Pruned BayesFuzzy 2 (PBF2), which is PBF incorporated with a feature selection technique based on Markov Blanket. With the incorporation of this technique, it's possible to deal with situations that contains a large amount of variables inside of the Markov Blanket of the class variable. The results show a loss in correct classification rate, that is already expected when we try to simplify further more the Markov Blanket. However, the availability to be able to deal with big scale problems is something to be considered.