Support and Confidence Based Rule Extraction Method for Neural Networks

Abstract

This paper presents a rule extraction method for competitive learning neural networks that are used for data clustering. First, a partition algorithm is used to divide attribute values into non-overlapped intervals. Consistency evaluation method adopted from rough set theory is used to partition attribute values. The generation of the set of adjoined intervals is controlled by the consistency evaluation against with the data distribution on the neural networks. By keeping the level of consistency, the set of adjoined intervals correctly reflects the data distribution on the networks. Second, instead of exhaustively traversing all combinations of the intervals to test possible rules, our method constructs the rules systematically and recursively from lower dimensions to higher ones. Using and adapting the techniques of evaluating amounts of support and confidence for an association rule, the constructed rules from our method are supported by the data clustering to the networks with adequate confidence. Finally, a rule reduction and merging algorithm is used to obtain a concise yet accurate set of rules. To verify the correctness of the constructed rules from our method, five benchmark problems are tested and results are compared. Comparison shows that the correctness of the rules generated from our method is more accurate than those from decision tree C4.5.