On the use of hard neural networks for symbolic learning application to company evaluation

The authors describe how the combination of two techniques can be used for symbolic learning: the hard neural network, built with an adaptation of the back-propagation algorithm, to capture the knowledge; and the decision tree generation ID3 algorithm to extract rules from the neural network. The resulting technique, developed in the framework of company financial health evaluation, produces good results in terms of rule base generality and readability. For the present, this method produced a more general and structured rule base than its counterpart obtained from direct application of the local-ID3 algorithm on the set of examples. In addition, the generation of intermediate features turned out to be of high interest.<>