Effectiveness of designing fuzzy rule-based classifiers from Pareto-optimal rules
Isao Kuwajima, Hisao Ishibuchi, Yusuke Nojima · 2008
In the field of data mining, two rule evaluation criteria called confidence and support are often used to evaluate a rule. Pareto-optimality of rules can be defined using these two criteria. The rules that are Pareto-optimal in the maximization of confidence and support are called Pareto-optimal rules. In this paper, we examine the effectiveness of designing fuzzy rule-based classifiers from Pareto-optimal rules and near Pareto-optimal rules. To show the effectiveness, we compare the Pareto-optimal (and near Pareto-optimal) rules with rules extracted by various rule evaluation criteria. In the design of classifiers, we use evolutionary multiobjective rule selection to obtain simple and accurate classifiers. Through computational experiments, we show that the best fuzzy rule with respect to each rule evaluation criterion is one of Pareto-optimal rules. We also show that fuzzy rule-based classifiers designed from Pareto-optimal rules have higher accuracy.