Mining Databases with Multiple Tables : Problems and Perspectives

We consider the problem of discovering patterns from a given logic that are significant (i.e. interesting and sufficiently valid) with respect to a given data set. We first define two types of patterns that extend the notions of query and clause, and we propose new measures of interest and confidence. In this framework, we establish connections between our measures, first-order logic and logics of probability of Halpern. Then, we present mining algorithms based on the generic levelwise search method proposed by Mannila, and discuss implementation issues in a relational database environment. Finally, we offer concluding remarks and suggestions for future research.