Constraint-based discovery and inductive queries: application to association rule mining

Abstract. Recently inductive databases (IDBs) have been proposed to afford the problem of knowledge discovery from huge databases. Querying these databases needs for primitives to: (1) select, manipulate and query data, (2) select, manipulate and query “interesting ” patterns (i.e., those patterns that satisfy certain constraints), and (3) cross over patterns and data (e.g., selecting the data in which some patterns hold). Designing such query languages is a long-term goal and only preliminary approaches have been studied, mainly for the association rule mining task. Starting from a discussion on the MINE RULE operator, we identify several open issues for the design of inductive databases dedicated to these descriptive rules. These issues concern not only the offered primitives but also the availability of efficient evaluation schemes. We emphasize the need for primitives that work on more or less condensed representations for the frequent itemsets, e.g., the (frequent) δ-free and closed itemsets. It is useful not only for optimizing single association rule mining queries but also for sophisticated post-processing and interactive rule mining.