Desenvolvimento de uma plataforma hÃbrida para descoberta de conhecimento em bases de dados

AUTOR(ES)
DATA DE PUBLICAÇÃO

2004

RESUMO

Artificial Neural Networks (ANN) have successfully been used in tasks as the mapping of complex functions and pattern recognition. This success is due to the ANN ability to make calculations of complicated and undetermined data, learn from examples, generalize the learned information, extract patterns and discover tendencies. Despite these advantages, it is generally not very easy to obtain explanations of how an ANN represents the solution of a problem. Due to this limitation, ANN have been considered inadequate to be used in applications of KDD (Knowledge Discovery in Databases) where the user wants to know the reasoning used by the network to obtain a conclusion. Intelligent Hybrid Systems (IHS) is an approach of Artificial Intelligence that has been used in resolution of problems where the application of one isolated technique is not sufficient to get satisfactory results. These systems are based on the integration of two or more intelligent techniques, aiming at overcoming the limitations of each technique. The dissemination of IHS has contributed to the emergence of Hybrid Neural Systems (HNS). The main research focus in HNS has been the integration of ANN, strongly data-based technique, with techniques that use symbolic representation, such as Fuzzy Logic and conventional symbolic algorithms. Neuro-Fuzzy Systems are an example of HNS that combine connectionist systems with fuzzy systems. In these systems, some rule extraction technique is applied to represent the knowledge embedded by the neural network in a comprehensible way. In addition to the rule extraction techniques of Neuro-Fuzzy Systems, several techniques for extraction of symbolic knowledge from other neural models have been proposed. The main goals of this work are investigating the paradigm of Neuro-Fuzzy Systems and the techniques for extraction of symbolic knowledge from ANN as an option to become ANN more adequate to the KDD process; and, as result of the investigation, modeling and developing a software tool, Neural Mining, based on the hybrid neural approach. The Neural Mining tool integrates, in an only environment, the neural model MLP (Multilayer Perceptron), the neuro-fuzzy models FWD (Feature-Weighted Detector) and FuNN (Fuzzy Neural Network), together with their rule extraction techniques, and the technique TREPAN (Trees Parroting Networks) that represents the knowledge embedded by an ANN in the form of a decision tree. The models and techniques are evaluated with respect to their generalization performance and comprehensibility of the extracted knowledge. In addition to the analysis in the data mining step and knowledge presentation step, two feature selection techniques are also investigated: the FWD technique and using the decision tree extracted by TREPAN. The experimental investigation is performed using a large scale credit assessment database of a real problem. As the results acquired demonstrate that the benefits obtained from the use of neurofuzzy models and techniques for extraction of symbolic knowledge from ANN are highly expressive, in the end of the investigation, considering the advantages of each model and technique, two hybrid neural solutions are proposed to the KDD process

ASSUNTO(S)

ciencia da computacao redes neurais artificiais (rna) intelligent hybrid systems (ihs) artificial intelligence sistemas neurais hÃbridos (snh) artificial neural networks (ann) hybrid neural systems (hns) sistemas hÃbridos inteligentes (shi) inteligÃncia artificial

ACESSO AO ARTIGO

Documentos Relacionados