On the Importance of Change of Representation in Induction Background: Success and Failure in Inductive Learning

The adequacy of current machine-learning techniques relies on an implicit built-in assumption in which simplicity and similarity play signiicant roles. This paper discusses on the insuuciency of these assumptions for domains where complex attribute interactions characterize the target concept (i.e. on diicult domains), and revives a model of induction where change of representation, through the construction of increasingly more abstract knowledge structures, contributes to increased learning performance. We describe two learning systems proved eeective on diicult domains, their success attributed to the use of knowledge structures that bridge the gap between primitive and complex representations. Our conclusions emphasize the importance of change of representation as a fundamental step during inductive learning.