Learning Relational Concepts at Different Levels of Granularity

In this paper, an alternative approach to the induction of relational concepts is presented. The underlying framework relies on the concept of exception, an exception being a counterexample left within the scope of a description devoted to classifying examples of the given target concept. While trying to characterize the target concept, first an initial description is searched for. Such a solution must be complete, although not necessarily consistent. This means that some counterexamples are allowed to be misclassified. As counterexamples (i.e., exceptions) must be taken into account in order to properly classify them, the corresponding learning process is performed in several steps, each step devoted to coping with exceptions generated during the previous one. Eventually, the process comes to an end, usually leading to a description that uses a kind of Vere’s counterfactuals to refine, at different levels of granularity, the underlying concept.