Hvbrid rule-extraction from support vector machines

Rule-extraction from artificial neural networks (ANNs) as well as support vector machines (SVMs) provide explanations for the decisions made by these systems. This explanation capability is very important in applications such as medical diagnosis. Over the last decade, a multitude of algorithms for rule-extraction from ANNs have been developed. However, rule-extraction from SVMs is not widely available yet. In this paper, a hybrid approach for rule-extraction from SVMs is outlined. This approach has two basic components: (1) data reduction using a logistic regression model and (2) learning based rule-extraction. The quality of the extracted rules is then evaluated in terms of fidelity, accuracy, consistency and comprehensibitity. The rules are also verified against the available knowledge from the domain problem (diabetes) to assure correctness and validity.