Boosted Rule Learner and its Properties
Mariusz Kubus · 2018
This chapter describes some properties of the boosted rule learner. It examines whether boosting improves the stability of a classifier in this form, and whether this method is robust to irrelevant variables. The chapter compares the classification accuracy of two algorithms that have implemented boosting: AdaBoost.M1, which builds an ensemble of trees, and SLIPPER, which builds a set of weighted rules. Cohen and Singer adapted boosting for inducing the rules in the SLIPPER algorithm. The SLIPPER algorithm was proposed for the binary classification problem. SLIPPER outperforms its predecessor RIPPER and even single classification trees CART due to accuracy of classification. One of the most advanced algorithms that follows the separate-and-conquer scheme is RIPPER. Its effectiveness is a result of a quite sophisticated pruning technic, which combines pre- and post-pruning, and a use of stopping criteria based on the minimum description length (MDL) principle.