Design of Fuzzy Rule-based Classifiers with Inhomogeneous Fuzzy Partitions

Abstract

Discretization of continuous attributes is a key issue in classifier design from numerical data. In the fuzzy systems community, generating membership functions from numerical data has been an important research topic. Fuzzy genetics-based machine learning (GBML), which is one of the frequently-used techniques to design fuzzy rule-based classifiers, has often used uniform fuzzy partitions to generate initial membership functions. In this paper, we apply a class entropy measure to the discritization of numerical attiributes in order to generate inhomogeneous interval partitions. Nonuniform asymmetric membership functions are constructed from the generated interval partitions by introducing a parameter called "fuzzification grade" (0: interval, 1: full fuzzification). Through computational experiments, we examine the effects of the fuzzification grade on the accuracy of fuzzy rule-based classifiers. We also propose an ensemble classifier which has several fuzzy rule-based classifiers with different fuzzification grades.