Our goal is to model the way people induce knowledge from rare and sparse data. This paper describes a theoretical framework for inducing knowledge from these incomplete data described with conceptual graphs. The induction engine is based on a non-supervised algorithm named default clustering which uses the concept of stereotype and the new notion of default subsumption, the latter being inspired by the default logic theory. A validation using artificial data sets and an application concerning an historic case are given at the end of the paper.