Classifying uncovered examples by Rule Stretching

Abstract

This paper is concerned with how to classify examples that are not covered by any rule in an unordered hypothesis. Instead of assigning the majority class to the uncovered examples, which is the standard method, a novel method is presented that minimally generalises the rules to include the uncovered examples. The new method, called Rule Stretching, has been evaluated on several domains (using the inductive logic programming system Virtual Predict for induction of the base hypothesis). The results show a significant improvement over the standard method.