dictive accuracy. Quinlan's ID3, an early decision
tree learner, initially used the information gain split
method. But Quinlan discovered that information
gain showed unfair favoritism toward attributes with
many outcomes. Consequently, gain ratio later became
the default split method.Though c4.5, a descendent of ID3, still defaults to
the gain ratio criterion, the user has the option to
request information gain. [4, page 86] Unfortunately,
it is unclear how switching to information gain aects
the learner's bias.One corollary of Cullen Schaer's Conservation
Law of Generalization Performance gives researchers
a better understanding of an issue concerning the
comparison of learners under certain situations. A
concept learning situation is a triple