Chinese Journal of Computers   Full Text
Title	Classification Rule Extraction Based on Fuzzy Area Distribution and Classification Reasoning Algorithm
Authors	LI Jie1),2) DENG Yi-Ming2) SHEN Shi-Tuan2)
Address	1)(Institute of Communication Standards Research, China Academy of Telecommunication Research of MII, Beijing 100045) 2)(School of Electronics and Information Engineering, Beijing University of Beihang, Beijing 100083)
Year	2008
Issue	No.6(934—941)
Abstract &   Background	Abstract A definition of rule relative compatibility is proposed based on membership degree distribution on the fuzzy area corresponding to every rule. Compared with the class compatibility, rule relative compatibility is able to show more information about the samples distribution difference among every fuzzy area. The Fuzzy Area Distribution Matrix is designed, by which the rule relative compatibility and class compatibility are calculated respectively. Moreover, the algorithm of classification rule extraction is put forward by rule relative compatibility. Different from the class compatibility approach, the one based on rule relative compatibility contains the membership degree distribution comparison of every rule. Furthermore, this compatibility is weighted by the relative amount of every class samples so that it can take the consideration of the global density dominance as well as the local quantity dominance for study space. In addition, the classification reasoning based on fuzzy rule is implemented by the defuzzifier algorithm. The procedure is better than the previous algorithm due to its interpretability and simplicity. In the end, Iris data and Wine data are used to validate the proposed algorithm of fuzzy classification rule extraction. The testing results prove that whether the sample is distributed homogeneously or not, the rule extraction approach based on rule relative compatibility attains higher classification rate. Keywords classification rule; classification reasoning; defuzzifier; fuzzy area distribution Background This work is supported by the Defense Pre-Research Project of the 11th Five-Year-Plan of China. As a branch of Artificial Intelligence, Pattern Classification is a sort of process of human thinking and reasoning by using computer. Based on fuzzy set and fuzzy proposition, Pattern Classification is just like the human reasoning due to their fuzzy characteristic. Furthermore, the reasoning process of Pattern Classification is comprehensive because of fuzzy rule’s interpretability. As one of the most promising research domains of fuzzy classification, initial fuzzy rule base generation has been widely studied recently. Neural Network has been applied for extraction of fuzzy rules, which consists of four layers and optimizes the membership function by employing pruning algorithm. Since the Neural Network has the capability of knowledge storage and reasoning, it is inefficient to append a step of rules generation for Neural Network. Kernel Machine has been used to fuzzy classification rule extraction, by which the rule base is generating from case space based on Support Vector Machine and then the Positive Definite Fuzzy Classifier is constructed. The Kernel Function Parameter, however, does depend on different applications, which is hard to be implemented by now. The Clustering algorithm is also applied for fuzzy rule generation to reduce the amount of rule base and the amount of rule attribute. While the rule base is simplified, the complexity is increased. The Genetic Algorithm is also a method to generate fuzzy rule, of which performance is determined by initial rule-base though. This research aims at the key techniques of fault classification for automatic fault diagnosis, which carries out the comprehensive and efficient classification rule base generation and classification reasoning by fuzzy theory. To guarantee both performance of classification and interpretability of reasoning, authors of this paper propose fuzzy rule extraction method using the rule relative compatibility defined in this paper. Rule Relative Compatibility considers both the relative amount of every class’s case and the global distribution in the case space. The Fuzzy Area Distribution Matrix is designed to calculate all sorts of compatibility, and then the algorithm of the initial complete rule base generation is designed based on this matrix. Also, based on the defuzzifier method, a classification reasoning approach is designed to simplify the procedure of fuzzy classification. The experiment results form the benchmark data demonstrates that the algorithm proposed by this contribution is able to obtain more good classification performance for both homogeneous cases and heterogeneous cases.