Effects of constructing fuzzy discretization from crisp discretization for rule-based classifiers

Crisp discretization is one of the most widely used methods for handling continuous attributes. In crisp discretization, each attribute is split into several intervals and handled as discrete numbers. Although crisp discretization is a convenient tool, it is not appropriate in some situations (e.g., when there is no clear boundary and we cannot set a clear threshold). To address such a problem, several discretizations with fuzzy sets have been proposed. In this paper we examine the effect of fuzzy discretization derived from crisp discretization. The fuzziness of fuzzy discretization is controlled by a fuzzification grade F. We examine two procedures for the setting of F. In one procedure, we set F beforehand and do not change it through training rule-based classifiers. In the other procedure, first we set F and then change it after training. Through computational experiments, we show that the accuracy of rule-based classifiers is improved by an appropriate setting of the grade of fuzzification. Moreover, we show that increasing the grade of fuzzification after training classifiers can often improve generalization ability.