Knowledge Discovery in Real Databases: A Report on the IJCAI-89 Workshop

able databases far outstrips the growth of corresponding knowledge. This creates both a need and an opportunity for extracting knowledge from databases. Many recent results have been reported on extracting different kinds of knowledge from databases, including diagnostic rules, drug side effects, classes of stars, rules for expert systems, and rules for semantic query optimization. The importance of this topic is now recognized by leading researchers. Michie predicts that “The next area that is going to explode is the use of machine learning tools as a component of large scale data analysis’’ (AI Week, March 15, 1990). At a recent NSF invitational workshop on the future of database research (Lagunita, CA, February 1990), “knowledge mining” was among the top five research topics. The viability of knowledge discovery is also being recognized by business and government organizations. American Airlines is looking for patterns in its frequent flyer databases. Banks are analyzing credit data to determine better rules for credit assessment and bankruptcy prediction. General Motors is automatically constructing diagnostic expert systems from a database of car trouble symptoms and problems found. The IRS is looking for patterns of tax cheating in its databases. Those are only some of the examples. Some of the research has matured enough to find its way into commercial systems for rule discovery in databases. Several such systems have appeared recently, including IXLTM, BEAGLETM, KnowledgeSeekerTM, AIMTM, and KnowledgeMakerTM. Knowledge discovery in databases is an interesting topic, drawing from several fields including expert systems, machine learning, intelligent databases, knowledge acquisition, case-based reasoning and statistics. The Knowledge Discovery in Databases workshop, held on August 20, 1989 in Detroit, MI, during IJCAI-89, had succeeded in bringing together many leading researchers in Machine Learning, Expert Databases, Knowledge Acquisition, Fuzzy Sets, and other areas. The workshop had interesting presentations and lively panel discussion, with lots of interaction. It helped to remove some of the misconceptions that Machine Learning researchers have about databases— i.e. databases are not static tables of simple data, but complex entities with transactions, security, and updates. While those researchers just want to use the data, the database people want to integrate the acquired knowledge into the database system. The workshop also helped to educate the database researchers about the available wealth of approaches to machine discovery. I was the chairman of the workshop. The program committee consisted of Jaime Carbonell, Carnegie Mellon University, William Frawley, GTE Laboratories, Kamran Parsaye, IntelligenceWare, Los Angeles, J. Ross Quinlan, University of Sydney, Michael Siegel, Boston University and MIT, and Ramasamy Uthurusamy, GM Research Laboratories. The workshop generated a significant international interest. We received 69 submissions from 12 countries: USA (39), Canada (9), UK (3), P.R. China (3), Italy (3), France (2), Sweden (2), India (2), Belgium (2), Germany (2), Japan (1), and Australia (1). Thirty nine contributors from 9 countries were invited to attend the workshop. The workshop was also attended by Robert Simpson from DARPA and Y.T. Chien, Director of AI & Robotics at the NSF. Nine excellent papers were presented in three sessions: Data-Driven Discovery Methods, KnowledgeBased Approaches, and Systems and Applications. The revised versions of the workshop papers will be included in the forthcoming collection on Knowledge Discovery in Databases, to be published by AAAI and MIT press in early 1991, and will not be discussed here. The workshop concluded with a stimulating panel discussion by Pat Langley, Larry Kerschberg and J. Ross Quinlan. There was general agreement that Knowledge Discovery is a promising research direction that will become more important as the number of domains for which there are no human experts increases. Applications to large real databases will require algorithms that are efficient and handle uncertainty well. More complex domains demand the use of more expressive (i.e., first-order) languages.