Instance-based method to extract rules from neural networks

It has been shown that a neural network is better than induction tree applications in modeling complex relations of input attributes in sample data. Those relations as a set of linear classifiers can be obtained from neural network modeling based on back-propagation. A linear classifier is derived from a linear combination of input attributes and neuron weights in the first hidden layer of neural networks. Training data are projected onto the set of linear classifier hyperplanes and then information gain measure is applied to the data. We propose that this can reduce computational complexity to extract rules from neural networks. As a result, concise rules can be extracted from neural networks to support data with input variable relations over continuous-valued attributes.