Complexity of Rule Sets Induced by Two Versions of the MLEM2 Rule Induction Algorithm

We compare two versions of the MLEM2 rule induction algorithm in terms of complexity of rule sets, measured by the number of rules and total number of conditions. All data sets used for our experiments are incomplete, with many missing attribute values, interpreted as lost values, attribute-concept values and do not care conditions. In our previous research we compared the same two versions of MLEM2, called true and emulated, with regard to an error rate computed by ten-fold cross validation. Our conclusion was that the two versions of MLEM2 do not differ much, and there exists some evidence that lost values are the best. In this research our main objective is to compare both versions of MLEM2 in terms of complexity of rule sets. The smaller rule sets the better. Our conclusion is again that both versions do not differ much. Our secondary objective is to compare three interpretations of missing attribute values. From the complexity point of view, lost values are the worst.