M.: An evolutionary algorithm for cost-sensitive decision rule learning

Abstract. Most of classication learning methods aim at the reduc-tion of the number of errors. However, in many real-life applications it is misclassication cost, which should be minimized. In the paper we propose a new method for cost-sensitive learning of decision rules from datasets. Our approach consists in modifying the existing system EDRL-MD (Evolutionary Decision Rule Learner with Multivariate Discretiza-tion). EDRL-MD learns decision rules using an evolutionary algorithm (EA). We propose a new tness function, which allows the algorithm to minimize misclassication cost rather than the number of classication errors. The remaining components of EA i.e., the representation of solu-tions and the genetic search operators are not changed. The performance of our method is compared to that of C5.0 learning system. The results show, that the modied EDRL-MD is able to eectively process datasets with non-equal error costs.