A rule extraction algorithm that scales between fidelity and comprehensibility

Abstract

Fidelity and comprehensibility are the common measures used in the evaluation of rules extracted from neural networks. However, these two measures are found to be inverse relations of one another. Since the needs of comprehensibility or fidelity may vary depending on the user or application, this paper presented a significance based rule extraction algorithm that allows a user set parameter to scale between the desired degree of fidelity and comprehensibility of the rules extracted. A detailed explanation and example application of this algorithm is presented as well as experimental results on several neural network ensembles. © 2012 Asian Network for Scientific Information.