Rule Extraction by Seeing Through the Model

Much effort has been spent during recent years to develop techniques for rule extraction from opaque models, typically trained neural networks. A rule extraction technique could use different strategies for the extraction phase, either a local or a global strategy. The main contribution of this paper is the suggestion of a novel rule extraction method, called Cluster and See Through (CaST), based on the global strategy. CaST uses parts of the well-known RX algorithm, which is based on the local strategy, but in a slightly modified way. The novel method is evaluated against RX and is shown to get as good as or better results on all problems evaluated, with much more compact rules.