Obtaining accurate classifiers with Pareto-optimal and near Pareto-optimal rules

In the field of data mining, confidence and support are often used to measure the quality of a rule. Pareto-optimal rules, which are Pareto-optimal in terms of confidence and support maximization, have an interesting characteristic that Pareto-optimal rules maximize other various rule evaluation criteria. In this paper, we examine the effectiveness of designing classifiers from Pareto-optimal rules. We consider not only Pareto-optimal rules but also near Pareto-optimal rules. To show the effectiveness, we compare classifiers obtained from Pareto-optimal and near Pareto-optimal rules with classifiers obtained from the rules that have large value in terms of other different rule evaluation criteria. Eight criteria are examined in this paper: CF, confidence, cover, Laplace, lift, random, slave, support. Through computational experiments, we show that classifiers obtained from Pareto-optimal rules have higher accuracy than those from rules extracted according to the other criteria. This work was presented in part and awarded as Young Author Award at the 13th International Symposium on Artificial Life and Robotics, Oita, Japan, January 31–February 2, 2008