Richard Segal, Oren Etzioni
A decision list is an ordered list of conjunctive rules (Rivest 1987). Inductive algorithms such as AQ and CN2 learn decision lists incrementally, one rule at a time. Such algorithms face the rule overlap problem - the classification accuracy of the decision list depends on the overlap between the learned rules. Thus, even though the rules are learned in isolation, they can only be evaluated in concert. Existing algorithms solve this problem by adopting a greedy, iterative structure. Once a rule is learned, the training examples that match the rule are removed from the training set. We propose a novel solution to the problem: composing decision lists from homogeneous rules, rules whose classification accuracy does not change with their position in the decision list. We prove that the problem of finding a maximally accurate decision list can be reduced to the problem of finding maximally accurate homogeneous rules. We report on the performance of our algorithm on data sets from the UC1 repository and on the MONK’s problems.