Top-down induction of reduced ordered decision diagrams from neural networks

Neural networks offer good generalization performance, noise robustness, and model complexity control. However, neural network mappings are expressed in terms of complicated mathematical functions that are inherently hard to understand. To overcome this limitation rule extraction methods have been proposed. This paper presents a novel method of rule extraction which recursively, in a top-down manner, builds a Reduced Ordered Decision Diagram. The diagram structure allows sharing of nodes, which partially overcomes two problems present in Decision Tree-based rule extraction - the problem of subtree replication and of training set fragmentation. A method for reducing the rule search space by identifying regions in which the network shows similar behavior is presented. Preliminary results of the method performance are reported.