Very bound to the logic of first-rate predicates, the formalism of con
ceptual graphs constitutes a knowledge representation language. The ab
straction of systems presents several advantages. It helps to render c
omplex systems more understandable, thus facilitating their analysis a
nd their conception. Our approach of conceptual graphs abstraction, or
conceptual clustering, is based on rectangular decomposition. It prod
uces a set of clusters representing similarities between subsets of ob
jects to be abstracted, organized into a hierarchy of classes: the Kno
wledge Space. Some conceptual clustering methods already exist. Our ap
proach is distinguishable from other approaches in as far as it allows
a gain in space and time. (C) 1998 Elsevier Science Inc. All rights r
eserved.