The influence of class order on rule induction results

The basic sequential rule induction algorithm does not define a specific class order. Class order is generally selected based on heuristics. We have implemented the basic rule induction algorithm together with several discretization and evaluation methods and varying class orders. Tests in several basic data sets show that the class order considerably influences the resulting accuracy of sequential rule induction. While the class order sometimes has a higher influence than the tested rule quality measures (FOIL and Likelihood Ratio) and the simple discretization algorithms, there is no single class order that provides best results for all tested combinations of discretization and rule quality measure approaches in all cases.