MLEM2 Rule Induction Algorithm with Multiple Scanning Discretization

In this paper we show experimental results on the MLEM2 rule induction algorithm and the Multiple Scanning discretization algorithm. The MLEM2 algorithm of rule induction has its own mechanisms to handle missing attribute values and numerical data. We compare, in terms of an error rate, two setups: MLEM2 used for rule induction directly from incomplete and numerical data and MLEM2 inducing rule sets from data sets previously discretized by Multiple Scanning and then converted to be incomplete. In both setups certain and possible rule sets were induced. For certain rule sets, the former setup was more successful for two data sets, while the latter setup was more successful for four data sets, for eight data sets the difference was not significant (Wilcoxon test, 5% significance level). Similarly, for possible rule sets the former setup was more successful for two data sets, while the latter setup was more successful for three data sets. Thus we may conclude that there is not significant difference between both setups and that we may use MLEM2 for rule induction directly from incomplete and numerical data.