AbstractMany artificial intelligence systems implicitly use notions of granularity in reasoning, but there is very little research into granularity itself. An exception is the work of Hobbs, who outlines several characteristics of granularity. In this paper, we describe an approach to representing granularity which formalizes in computational terms most of Hobbs' notions, often refining and extending them. In particular, two types of granularity have been delineated: aggregation and abstractin. Objects can be described at various grain sizes and connected together into a granularity hierarchy which allows focus shifts along either aggregation or abstractiondimension.Granularity hierarchies can be used in recognition. An especially good domain for granularity-based recognition is educational diagnosis. In an intelligent tutoring system, the ability to recognize student behaviour at varying grain sizes is important both for pedagogical reasons (in order to respond to the student at various levels of detail) and for reasons of robustness in diagnosis (obscure student behaviour can be recognized at least at a coarse grain size). We briefly discuss how we have used granularity hierarchies in the recognition of novice LISP programming strategies, and show how this enhances recognition and leads toward planning appropriate feedback for the student.