We argue that in a distributed context, such as the Semantic Web, ontology engineers and data creators often cannot control (or even imagine) the possible uses their data or ontologies might have. Therefore ontologies are unlikely to identify every useful or interesting classification possible in a problem domain, for example these might be of a personalised nature and only appropriate for a certain user in a certain context, or they might be of a different granularity than the initial scope of the ontology. We argue that machine learning techniques will be essential within the Semantic Web context to allow these unspecified classifications to be identified. In this paper we explore the application of machine learning methods to FOAF, highlighting the challenges posed by the characteristics of such data. Specifically, we use clustering to identify classes of people and inductive logic programming (ILP) to learn descriptions of these groups. We argue that these descriptions constitute re-usable, first class knowledge that is neither explicitly stated nor deducible from the input data. These new descriptions can be represented as simple OWL class restrictions or more sophisticated descriptions using SWRL. These are then suitable either for incorporation into future versions of ontologies or for on-the-fly use for personalisation tasks.