Linguistic rule extraction from neural networks for descriptive data mining

Abstract

There are two main goals of knowledge discovery: prediction and description. Description deals with identifying patterns for the purpose of presenting them to users in a form understandable by humans. The ability of neural networks to learn patterns from noisy data made them a popular tool for data mining. The problem is, however, that neural networks do not provide description of the patterns they discover. In knowledge discovery for decision making the comprehensibility of discovered patterns is sometimes more important than their predictive capability. We describe a framework for a neural network based data mining system which presents discovered patterns in a comprehensible form. In this framework a neural network is first trained on a set of training data and a rule extraction technique is then applied in order to extract explicit knowledge from the network and represent it in the form of crisp and fuzzy If-Then rules. The framework is illustrated with an application.