Metaqueries for Data Mining

This chapter presents a framework that uses metaqueries to integrate inductive learning methods with deductive database technologies in the context of knowledge discovery from databases. Metaqueries are second-order predicates or templates, and are used for (1) Guiding deductive data collection, (2) Focusing attention for inductive learning, and (3) Assisting human analysts in the discovery loop. We describe in detail a system that uses this idea to unify a Bayesian Data Cluster with the Logical Data Language (LDL++), and show the results of three case studies, namely, discovering regularities from a knowledge base, discovering patterns and errors from a large telecommunication database, and discovering patterns and errors from a large chemical database. The patterns discovered using metaqueries are implication rules with probabilities. These rules can link information from many tables in databases, and they can be stored persistently for multiple purposes, including error detection, integrity constraints, or generation of more complex metaqueries. 15.1 Introduction Recent progress in knowledge discovery from databases (Cercone and Tsuchiya 1993; Piatetsky-Shapiro 1993) has shown that inductive hypothesis generation, deductive hypothesis veriication, and human intituation, are all crucial components for an eeective discovery system. Inductive learning is essential for generating hypotheses from data automatically, deductive database technology is a natural tool for gathering evidence in support of existing hypotheses, while human intuition (which may be inspired by the results of machine discovery) is necessary for generating and selecting the most promising hypotheses in a practical manner. However, the integration of these three components