Inducing Modular Rules for Classification

Abstract

This chapter begins by considering a method of post-pruning decision rules generated via a decision tree, which has the property that the pruned rules will not generally fit together to form a tree. Rules of this kind are known as modular rules. When using modular rules to classify unseen test data a conflict resolution strategy is needed and several possibilities for this are discussed. The use of a decision tree as an intermediate representation for rules is identified as a source of overfitting.