Rule extraction from structured neural network for pattern classification

Artificial neural networks are now widely used for various pattern classification problems and possess better accuracy rates than traditional statistical classifiers. But they are often regarded as black boxes as the interpretation of their inner working is quite difficult. Researchers have been developing algorithms that extract rules from neural network models and can allow one to explain the decision process of the network. Sparse architectures and structured models are easy to analyse. In this work an algorithm for extraction of rules from an artificial neural model for solving pattern classification problems, capable of handling real life vague information expressed by linguistic variables, has been devised. A sparse structured architecture of fractally connected feedforward multilayered neural networks has been used and its efficiency in rule extraction compared to the fully connected network has been studied. The ease of rule extraction by the proposed sparse fractal network compared to full connection network has been proved by simulation with two data sets