Rule Extraction Based on Data Dimensionality Reduction Using RBF Neural Networks

摘要

Compact rules is desirable in the task of rule extraction. Since there are often redundant or irrelevant attributes in data sets, removing the redundant or irrelevant attributes from the data sets can lead to more compact rules. In this paper, firstly, a novel method, a separability-correlation measure (SCM), is used to rank the importance of attributes, and an RBF classifier is used to evaluate the best subset of attributes to be retained. Secondly, large overlaps between clusters of the same class are allowed in order to reduce the number of hidden units in the RBF network. Thirdly, rule extraction is carried out based on the retained subset of attributes used as the input to the RBF neural network. Simulations show that this procedure lead to more compact rules.