Coupling Two Complementary Knowledge Discovery Systems

Lawrence B. Holder and Diane J. Cook University of Texas at Arlington Box 19015, Arlington, TX 76019-0015 Email: fholder,[email protected] Abstract Most approaches to knowledge discovery concentrate on either an attribute-value representation or a structural data representation. The discovery systems for these two representations are typically di erent, and their integration is non-trivial. We investigate a simpler integration of the two systems by coupling the two approaches. Our method rst executes the structural discovery system on the data, and then uses these results to augment or compress the data before being input to the attribute-value-based system. We demonstrate this strategy using the AutoClass attribute-valuebased clustering system and the Subdue structural discovery system. The results of the demonstration show that coupling the two systems allows the discovery of knowledge imperceptible to either system alone. Introduction With the increasing amount and complexity of today's data, there is an urgent need to improve the discovery of knowledge in large databases. Numerous approaches have been developed for discovering knowledge in databases using a linear, attribute-value representation. Although much of the data collected today has an explicit or implicit structural component (e.g., spatial or temporal), few discovery systems are designed to handle this type of data (Fayyad, Piatetsky-Shapiro, & Smyth 1996). One reported method for dealing speci cally with structural data is with the Subdue system (Cook, Holder, & Djoko 1996). Subdue provides a method for discovering substructures in structural databases using the minimum description length (MDL) principle introduced by Rissanen. Although numerous approaches exist for either attribute-value-based or structural discovery, no system provides a general approach to performing discovery in the presence of both types of data. An alternative to a uni ed approach is to separate the attribute-value and structural components of the Copyright 1998, American Association for Arti cial Intelligence (www.aaai.org). All rights reserved. 1 2 1 2 2 1 4 3 1 2 5 3