The Fuzzy Classifier Based On AFS Theory

AFS (Axiomatic Fuzzy Set) theory was proposed first as a new analysis method of fuzzy mathematics. In the framework of AFS theory, the membership functions and their logic operations for fuzzy concepts can be impersonally determined according to original data and facts. AFS theory has been applied to data mining, pattern recognition and failure diagnosis.Professor Liu proposed a fuzzy classification algorithm based on AFS theory. The classifier designed by this algorithm is able to obtain the description for every class and gives the right label of class to the unknown data at a high rate. However, the experiment results demonstrate that the description for the class is too complicated to understand. This article gives a new method based on the original algorithm to construct a classifier. We aggregate the fuzzy concepts using the AFS fuzzy logic operators. We use the method of decision tree in the process of aggregation, which makes the aggregation more efficient and faster. This article uses the criterions proposed by Professor Liu, which distinguish the intra-class data from inter-class data by the membership degrees, to select the aggregated fuzzy concepts. Then we use the selected fuzzy concepts to generate fuzzy rules. In the process of testing the unknown data, we consider the classification information generated by all the fuzzy rules, and the classification result is given after weighting all the classification information. Since we set a special depth of the decision trees during the aggregation of fuzzy concepts, the proposed algorithm can reduce the complexity of descriptions for every class. In the other way, we use the weighted information in the process of testing the unknown data (not dependent on a special single rule). Comparing with the original algorithm, we find that the proposed algorithm can reach a high accuracy rate of classification.We use the three datasets, Wine, Breast and Iris in our experiments, and compare the proposed algorithm with other classification algorithms including C4.5, KNN, Decision Table, JRip, NNge, OneR, PART and Ridor. The accuracy rate of classification is about 95% in the three experiments and better than most of other classification algorithms.