Qualitative Knowledge Discovery

Knowledge discovery and data mining deal with the task of finding useful information and especially rules in unstructured data. Most knowledge discovery approaches associate conditional probabilities to discovered rules in order to specify their strength. In this paper, we propose a qualitative approach to knowledge discovery. We do so by abstracting from actual probabilities to qualitative information and in particular, by developing a method for the computation of an ordinal conditional function from a possibly noisy probability distribution. The link between structural and numerical knowledge is established by a powerful algebraic theory of conditionals. By applying this theory, we develop an algorithm that computes sets of default rules from the qualitative abstraction of the input distribution. In particular, we show how sparse information can be dealt with appropriately in our framework. By making use of the duality between inductive reasoning and knowledge discovery within the algebraic theory of conditionals, we can ensure that the discovered rules can be considered as being most informative in a strict, formal sense. draft – 2008-07-20