4.10 Further Reading 143 of its instances is positive. Instead, making each instance in a bag contribute equally to its label is the key element that allows standard learning algorithms to be applied. Otherwise, it is necessary to try to identify the "special" instances that are the key to determining the bag's label. 4.10 FURTHERREADING The 1R scheme was proposed and thoroughly investigated by Holte (1993). It was never really intended as a machine learning "method." The point was more to demonstrate that very simple structures underlie most of the practical datasets being used to evaluate machine learning schemes at the time and that putting high-powered inductive inference schemes to work on simple datasets was like using a sledgehammer to crack a nut. Why grapple with a complex decision tree when a simple rule will do? The scheme that generates one simple rule per class is due to Lucio de Souza Coelho of Brazil and Len Trigg of New Zealand, and it has been dubbed hyperpipes. A very simple algorithm, it has the advantage of being extremely fast and is quite feasible even with an enormous number of