Making knowledge extraction and reasoning closer

The paper shows how a logic-based database language can support the various steps of the KDD process by providing a high degree of expressiveness, and the separation of concerns between the specification level and the mapping to the underlying databases and data mining tools. In particular, the mechanism of user-defined aggregates provided in LDL++ allows to specify data mining tasks and to formalize the mining results in a uniform way. We show how the mechanism applies to the concept of Inductive Databases, proposed in [2,12]. We concentrate on bayesian classification and show how user defined aggregates allow to specify the mining evaluation functions and the returned patterns. The resulting formalism provides a flexible way to customize, tune and reason on both the evaluation functions and the extracted knowledge.