The Fuzzy Rules Based Classification Algorithm

Classification is a very critical topic in data mining,and it has a wide range of application in the field of pattern recognition,statistics,machine learning,etc.The fuzzy rules based classification algorithm has many advantages,such as high classification accuracy,and the classification results are semantic,interpretable,and easily understood by the users,etc.Therefore,this thesis mainly studies two kinds of classification algorithms based on fuzzy rules,one is the construction of fuzzy rule classifier based on data,and the other is the construction of fuzzy oblique decision tree based on Axiomatic Fuzzy Sets(AFS).Main topics include:(1)An algorithm for constructing fuzzy rule classifier based on data is proposed in this chapter.Firstly,a forward greedy fast attribute reduction algorithm(NRS_FS_FAST)is used to preprocess the data.In the setting of neighborhood size,the threshold is set based on the threshold vector of attribute standard deviation.The above method can avoid some shortcomings.For example,single neighborhood size setting can cause the problem that the decision attribute is over-dependent on the intensive data,and then brings a large error to attribute reduction.Besides,in the simplification process of the rule base,this chapter uses the frequency of rules to calculate the confidence degree,which considers several samples and solves the shortcoming of the classifier instability caused by the traditional confidence algorithm that only one sample is taken into account.Finally,in order to test the superiority of the proposed algorithm,this chapter compares the classification accuracy and the number of rules with 6 traditional classification algorithms on 8 UCI data sets.Experimental results show that the average classification accuracy of the proposed classification algorithm is the highest,and the rules are very simple with good semantics and interpretability.(2)This chapter proposes a fuzzy oblique decision tree(FODT)based on AFS.Firstly,to avoid the influence of redundant attributes,the NRS_FS_FAST algorithm is used to preprocess the data.Secondly,the membership functions are generated automatically based on the AFS theory.They reduce the subjectivity of selecting membership function and overcome the problem in which the traditional decision trees have no semantic interpretation.Then,an efficient method(FREA)for extracting fuzzy rules is designed on account of fuzzy confidence,and the extracted fuzzy rules are then used to construct the FODT.In contrast with traditional decision trees in which only a single feature is taken into account at each none-leaf node,the growth of the FODT depends on fuzzy rules composed by a number of properties.In addition,the construction process of the FODT is affected by the threshold ?,so the genetic algorithm is used to optimize the parameter ?,which can ensure the classification accuracy and reduce the size of the tree as much as possible.Finally,the FODT algorithm is compared with the traditional decision trees(C4.5,LADtree,BFTree,Simple Cart and NBTree)and the FRDT algorithm on 9 UCI data sets.The experimental results show that the average classification accuracy of the FODT is the highest,and the average number of rules is less than that of traditional decision trees.And the Holm test shows that the FODT is superior to the traditional decision trees on the classification accuracy.Finally,the two classification algorithms are summarized,and the research direction and content of the research are prospected.