We describe an approach for conceptual user modeling as realized in the
OySTER meta web search engine. Instead of collaboratively modeling user
interests we use web document classifications in order to describe
individual user models. Information relevance is expressed with respect to
an underlying ontology of text categories. Logic expressions over
semi-lattices are interpreted as horn clauses---thus allowing to prove
different levels of interestingness. Furthermore, this approach presents a
well-defined learning problem for inductive logic programming which yields
inspectable user models that include sets of interest aspects. By using both
explicit positive and negative feedback for both interest and explicit
dis-interest we can use few examples to generate a larger set of labeled
data for the learning task.