Changing representation of learning examples while inducing classifiers based on decision rules

Decision rules induced from examples are used to predict classification of new objects. Improving classification accuracy may be obtained by changing the original representation of the learning data. Two different approaches to such a transformation are considered in this paper: selecting the subset of the most relevant attributes with the wrapper approach, and modifying the presence of some learning examples in the learning set by the bagging technique. Both approaches are applied to the rule induction algorithm MODLEM and experimentally evaluated on several data sets.