An Evolutionary Algorithm Using Multivariate Discretization for Decision Rule Induction

We describe EDRL-MD, an evolutionary algorithm-based system, for learning decision rules from databases. The main novelty of our approach lies in dealing with continuous – valued attributes. Most of decision rule learners use univariate discretization methods, which search for threshold values for one attribute at the same time. In contrast to them, EDRL-MD simultaneously searches for threshold values for all continuous-valued attributes, when inducing decision rules. We call this approach multivariate discretization. Since multivariate discretization is able to capture interdependencies between attributes it may improve the accuracy of obtained rules. The evolutionary algorithm uses problem specific operators and variable-length chromosomes, which allows it to search for complete rulesets rather than single rules. The preliminary results of the experiments on some real-life datasets are presented.