Articles Building Intelligent Learning Database Systems

Xindong Wu

Induction and deduction are two opposite opera- cannot support any representation facility for tions in data-mining applications. Induction complex semantic information. extracts knowledge in the form of, say, rules or Traditional database technology can only decision trees from existing data, and deduction support facilities for processing data. Along applies induction results to interpret new data. An with the developments of other subjects, such intelligent learning database (ILDB) system inte- as decision science and AI, more and more grates machine-learning techniques with database applications need facilities for supporting both and knowledge base technology. It starts with existing database technology and performs both data management and knowledge manage- induction and deduction. The integration of data- ment (such as rules for automatic data infer- base technology, induction (from machine learn- ring and management of integrity constraints ing), and deduction (from knowledge-based sys- between data). tems) plays a key role in the construction of ILDB To widen the applicability of database tech- systems, as does the design of efficient induction nology to these new kinds of application, and deduction algorithms. This article presents a object-oriented approaches are currently pop- system structure for ILDB systems and discusses ular in processing structurally complex objects, practical issues for ILDB applications, such as and deductive databases or logic databases instance selection and structured induction. have been expected to support a solution to those applications where both knowledge and data models are needed. However, the knowl- ver the past 30 years, database research edge bases (which contain deductive rules has developed technologies that are and/or semantic information such as the con- Onow widely used in almost every com- ceptual hierarchy among data) in existing puting and scientific field. However, many new deductive database systems can only be built advanced applications, including computer- up manually with known technology. Auto- aided design (CAD) and computer-aided manu- matic knowledge acquisition or learning from facturing (CAM), have revealed that traditional databases directly in deductive database sys- database management systems (DBMSs) are tems has become a central and difficult prob- inadequate, especially in the following cases: lem in deductive database systems research. Conventional database technology has laid Existing work relevant to knowledge acqui- particular stress on dealing with large amounts sition falls into the following four categories: of persistent and highly structured data effi- (1) adding an induction engine to an existing ciently and using transactions for concurrency database system in an ad hoc way to imple- control and recovery. For some applications, ment rule induction from (or data mining in) such as CAD and CAM where the data schema- databases, (2) designing a specific engine to ta need to vary frequently, new data models are learn from a domain-specific data set, (3) needed. building ontologies and knowledge bases for In some applications, such as geographic expert systems, and (4) designing various data and image data, the semantic relation- learning algorithms that have no direct con- ships among data (such as the variations and nections with existing database technology. developments of real-world entities in func- However, when we integrate machine-learning tion, performance, structure, and status, with techniques into database systems to imple- time and external variables’ variations) need to ment data mining or knowledge acquisition be represented as well as the data itself. Con- from databases, we face many problems, such ventional data models in database technology as (1) efficient learning algorithms because

Copyright © 2000, American Association for Artificial Intelligence. All rights reserved. 0738-4602-2000 / $2.00 FALL 2000 61 Articles

realistic databases are typically large and noisy exchanges information with users in the form and (2) expressive representations for both of pull-down menus. KBMS and DBMS are facil- data (for example, tuples in relational databas- ities to support knowledge base and database es, which represent instances of a problem management functions. DB and KB denote domain) and knowledge (for example, rules in databases and knowledge bases, respectively, a rule-based system, which can be used to solve and OS indicates operating system facilities. users’ problems in the domain, and the seman- The knowledge-acquisition engine, K.A. Engine, tic information contained in the relational implements induction from databases. I/D Because schemata). Engine is an inference and deduction engine. Meanwhile, although some commercial suc- Utility contains a set of common procedures database cesses have been found in existing learning sys- that are shared by K.A. Engine, KBMS, and technology tems, there are limitations on current machine- DBMS. Access Storage Interface is composed of learning techniques for both research and the basic knowledge and data operators. has found applications. The limited industrial support to wide the machine-learning community is an exam- KBMS ple of such limitations. Because database tech- The KBMS module supports facilities for inter- applications nology has found wide applications in various actively building, adapting, and displaying in various fields, it will surely generate significant effect knowledge bases; checking for semantic incon- fields, it on machine-learning research if we can couple sistencies; sorting knowledge bases to imple- them well. Therefore, research on knowledge ment efficient chaining (Wu 1993a); and edit- will surely acquisition from databases can be viewed as an ing knowledge base files. generate important frontier for both database and machine-learning technology (Wu 1993c). DBMS significant This article discusses issues in integrating The DBMS module is based on a commercial effect on machine learning with database and knowl- relational DBMS. Users can do conventional edge base technology to construct intelligent database operations by simply calling the com- machine- learning database (ILDB) systems. In the fol- mercial system. However, a new function, List learning lowing section, I outline the system structure a Relation, is developed here to translate rela- research if we of an ILDB system that I have developed. In tional files into the Prolog-based representa- Induction from Databases and Deduction of tion (Wu 1993b) suitable for use by the K.A. can couple Induction Results, I survey existing induction Engine. This representation relates to the prob- them well. and deduction techniques, respectively, for lem of expressive representations for both data ILDB systems. I follow these sections with a dis- and knowledge. It binds actual relational data Therefore, cussion of some essential problems in practical and the data schema together in an explicit research on systems construction. way and can represent semantic information (such as logic implication and constraints knowledge between attributes or fields in entities) as well acquisition System Structure as all the information that can be represented from An ILDB system (Wu 1995) supports database in the widely adopted entity-relationship (E-R) and knowledge base management functions as model. databases can well as learning facilities. It provides mecha- K.A. Engine be viewed as nisms for (1) preparing and translating stan- dard (for example, relational) database infor- There are two submodules in the K.A. Engine: an important mation into a form suitable for use by its (1) semantic information, which generates frontier for induction engines, (2) using induction tech- semantic networks from relational database niques to extract knowledge from databases, schemata in an interactive manner, and (2) rule both database and (3) interpreting the knowledge produced induction, which constructs decision trees and and machine- to solve users’ problems. With an ILDB system, generates rules. learning one can, for example, produce 100 to 200 con- junctive rules for 50 diseases from 2 million I/D Engine technology medical cases of the 50 diseases. Then, the The I/D Engine tests the knowledge produced ILDB system can use the rules in two different by the K.A. Engine on new data sets and inter- ways: (1) keep these rules instead of the origi- prets it to solve users’ problems. nal cases because the original cases might take a large space and (2) use these rules to diagnose Induction from Databases new cases. Figure 1 shows the system structure of my Among the functions shown in figure 1, the KESHELL2 system (Wu 1995). In the diagram, K.A. Engine is the central module in an ILDB Monitor is a man-machine interface that system. The K.A. Engine relates to the design of

62 AI MAGAZINE Articles

Monitor

K.A.Engine

I/D Engine

KBMS DBMS

Utility

Access Storage Interface O S

KB DB

Figure 1. The System Structure of an ILDB System.

efficient learning algorithms, which can gener- generate a description, say, production rules or ally be divided into three categories:1 (1) super- a decision tree, that distinguishes instances of vised classification, (2) unsupervised cluster- each class from other classes. ing, and (3) association analysis. This section Attribute-based induction algorithms (such provides a review of existing techniques in as ID3 [Quinlan 1986], c4.5 [Quinlan 1993]; AQ, these categories, all of which can be integrated CN2 [Clark and Niblett 1989]; AE1 [Hong 1985], into the system structure in figure 1. and HCV [Wu 1993d]), incremental induction × × Let E = D1 … Da be a finite attribute space algorithms (such as ID5R [Utgoff 1989], and the of a dimensions, where each Dj (j = 1, …, a) is a version-space method [Mitchell 1977]) fall into finite set of symbolic values or a numeric inter- the supervised classification category. val. An instance e = (v1, …, va) is an element of Unsupervised clustering (or concept forma- E means vj Dj. Each instance in a classifica- tion [Langley 1987]) deals with the discovery of tion belongs to a known class that, say, has a new concepts from unclassified data. The data specific name in E. The classification task is to input for clustering is similar to that for classi-

FALL 2000 63 Articles

fication, but the significant difference is that existing database technology. When the avail- no class information is available for each able data sets are not well organized or contain instance. Well-known algorithms in unsuper- too much noise, both attribute-based induc- vised clustering are CLUSTER/2, UNIMEM, COBWEB tion and inductive logic programming perform (Fisher 1996), CLASSIT, AUTOCLASS, BIRCH (Zhang, poorly, although they can sometimes deal with Ramakrishnan, and Livny 1997), and the con- a small amount of noise. Connectionist and nectionist Kohonen self-organizing map and statistical methods are able to show their sig- backpropagation (Dayhoff 1990). nificance in these environments. When exam- Association analysis (Han, Pei, and Yin 2000; ples in a data set are not classified with distinc- Agrawal and Srikant 1994; Agrawal, Imielinksi, tive concepts, only connectionist and and Swami 1993) starts with a different data statistical clustering methods can be adopted

environment. Let I = {i1, i2, …, iN} be a set of N to carry out unsupervised learning. In practical distinct literals called items and D a set of ILDB systems, we would need to construct an transactions over I. Each transaction is a set of algorithm library containing learning algo- items i1, i2, …, ik I. An association rule is an rithms of different paradigms and correspond- implication of the form A → B, where A, B I, ing documentation for the user to refer to, and In practical and A B = . A is called the antecedent of the choose among, the algorithms. ILDB systems, rule, and B is called the consequent. we would A set of items (such as the antecedent or the consequent of a rule) is called an item set. The Deduction of Induction Results need to number of items in an item set is the length (or Because real-world databases are typically construct an size) of the item set. An item set of some length incomplete and noisy, induction results cannot k is referred to as a k–item set. Each item set has be assumed to be perfect. When induction algorithm an associated statistical measure called support, results take the form of rules, interpreting them library denoted as supp. For an item set A I, supp(A) to classify a new instance needs to face three containing = s, if the fraction of transactions in D contain- possible cases that demand different actions. ing A is equal to s. A rule A → B has a measure First is no match: No rules match the instance. learning of strength called confidence (denoted as conf), Second is single match: One or more rules indi- algorithms of which is defined as the ratio supp(A cate the same class match. Third is multiple B)/supp(A). The problem of association analy- match: More than one rule matches the different sis is to generate all rules A → B that have both instance and indicates different classes. paradigms support and confidence greater than or equal The third case does not apply to decision to some user-specified thresholds, called mini- trees produced by ID3-like algorithms, but and mum support (minsupp) and minimum confi- when the trees are decompiled into production corresponding dence (minconf), respectively. rules (Quinlan 1993), the production rules will Association analysis from large databases has face the same problems. documenta- received much attention recently. To discover a In the single-match case, the choice of class tion for the useful and interesting association analysis, a to the instance is naturally the class indicated user to refer wide range of problems have been investigated by the rules. Deduction-time processing deals over such diverse topics as generalized associa- mainly with the conflict resolution in the third to, and choose tion rules (Aggarwal and Yu 1998; Tsur et al. case and class estimation for the first case. Exist- among, the 1998; Agrawal and Srikant 1994; Agrawal, ing techniques for dealing with the first and Imielinksi, and Swami 1993), measurements of third cases are both exclusively based on prob- algorithms. interestingness (Aggarwal and Yu 1998), quan- ability estimation. Among them, the measure of titative association rules (Srikant and Agrawal fit for dealing with the no match case and the 1996), multiple-level association rules (Han estimate of probability for handling the multiple- and Fu 1999), and association with sequential match case developed in AQ15 (Michalski et al. patterns (Agrawal and Srikant 1995). 1986) have been adopted widely in knowledge Every learning paradigm has its own advan- discovery and data mining. tages and disadvantages. For example, attri- The measure-of-fit and estimate-of-probabil- bute-based induction algorithms cannot pro- ity methods perform well with problem duce first-order rules on their own. Inductive domains where no real-valued attributes are logic programming research can contribute in involved. However, when a problem contains this regard, although it is in general less effi- attributes that take values from continuous cient for problems where attribute-based domains (that is, real numbers or integers), induction works. However, because it requires their performance, especially in terms of accu- background knowledge before learning can be racy, decreases. In existing induction algo- carried out, inductive logic programming has rithms, dealing with continuous domains is not found much realistic application with based on the discretization of them into a cer-

64 AI MAGAZINE Articles tain number of intervals. There are quite a few pose a complex problem that might be very strategies available for discretization, such as large in size into a number of subproblems by the information gain–based methods using domain knowledge and apply an induc- (Dougherty, Kohavi, and Sahami 1995; Quin- tion algorithm to each of the subproblems. lan 1993; Fayyad and Irani 1992). Once each When induction results from all these subprob- continuous domain has been discretized into lems are put in a single knowledge base, we intervals, the intervals are treated as discrete need chaining mechanisms to perform deduc- values in induction and deduction. Discretiza- tion on examples from the complex problem. tion is the standard way existing induction sys- There has been some significant work in the tems have taken. However, discretization of design of efficient chaining algorithms for continuous domains does not always fit accu- expert systems (Wu 1993a), but dealing with rate interpretation. To say an age greater than no match and multiple match cases needs 50 is old or a temperature above 32 degrees more attention in the data-mining context. Centigrade is high is fuzzy. In such cases, fuzzy Constructive induction: None of the AQ- interpretation of the discretized intervals at like, ID3-like and HCV-like algorithms (Wu deduction time could be valuable. Rather than 1993c) need explicit, built-in background taking the cut points decided by a discretiza- knowledge, which is why they are sometimes tion method as sharp borders for intervals, we called empirical learning methods. Such learning can instead place some kind of curve at each methods are different in nature from the cut point as a fuzzy border. With these fuzzy knowledge-rich learning methods, such as AM borders, a value can be classified into a few dif- and EURISKO developed by Lenat (1983, 1979), ferent intervals at the same time, with varying explanation-based learning, and inductive log- degrees. Thus, a single-match case could ic programming. change to a multiple match, and a no-match However, there is always implicit back- case could change to a single or even multiple ground knowledge embedded in the formula- match. Deduction with fuzzy borders of dis- tion of solution spaces and in the representa- cretized intervals is called fuzzy matching. In tion of examples. When a solution space turns the multiple-match case, we can take the inter- out to be inadequate, representation modifica- val with the greatest degree as the value’s dis- tion is needed, and the modification process crete value. Wu (1999) describes an implemen- typically involves searching for useful new tation of the fuzzy matching techniques. descriptive features (constructive induction) in terms of existing features or attributes. AQ17 (Wnek and Michalski 1994) of the AQ-like fam- Practical Issues ily has been developed to implement iterative When building practical ILDB systems, we construction of new attributes based on exist- need to face the following important problems, ing ones. Zheng (1995) has also tried a method in addition to noise handling and dealing with called X-of-N attributes in constructive deci- both numeric and nominal data, which have sion-tree construction. received wide attention in the design of various Constructive learning has become a strong data-mining systems. theme in inductive learning research. One of Instance selection: Dealing with very large the difficulties in constructive learning is that databases is one of the defining challenges in the complexity in some cases (such as iterative data-mining research and development. No feature construction) is extreme, but there are matter how powerful computers are or will be situations in which it is a necessary part of in the future, data-mining researchers and learning. practitioners must consider how to manage Integration of object-oriented design to ever-growing data that can be too large (for learn semantic information from relational example, with terabytes of data) to be schemata involving complex objects: The processed at one time. Instance selection is about entity-relationship (E-R) model is one of the approaches that select or search for a portion of most successful methods of formulating useful a large database that can be used in data min- abstract models in the conceptual structure ing instead of the whole database. One of the design of databases and the key design aid for major approaches for instance selection is sam- conventional databases implemented under a pling, in which a sample is selected for testing wide variety of commercially available systems. and analysis. Other major approaches include By focusing on the entities and their relation- windowing, data reduction, and selection of ships, the E-R model structures the way design- representative instances. ers approach the problem of creating extensi- Structured induction: The basic idea of ble databases. However, there are two structured induction (Shapiro 1987) is to decom- substantial problems here. One is that trans-

FALL 2000 65 Articles

forming an E-R model into a relational model mining techniques with existing database tech- during the logical design of databases results in nology has been a popular topic for both com- the loss of some semantic information that munities. Knowledge discovery and data min- exists in the E-R model. In other words, the ing can augment the ability of existing entities and relationships are not distinguished database systems to represent, acquire, and in the relational data model. The other prob- process a collection of expertise such as those lem is that the relationship types of the E-R that form part of the semantics of many model are too simple to express complex advanced applications. In the meanwhile, semantic features of relationships among and there should be ways to further explore new within entities. It is impossible for the relation- information from data processing. For example al data model to describe the changes in rela- (Wu and Craske 1997), query results from data- tionship(s) and other entities caused by an bases with existing DBMSs could be a good entity in an E-R model. The integration of source of information for providing hypothe- object-oriented design is expected to provide ses for knowledge discovery and data mining. facilities for expressing and acquiring these kinds of semantic information. The semantic information can be used to guide and facilitate Conclusions knowledge discovery in ILDB systems. Knowledge acquisition from databases has Incremental induction in the case of been worked over by researchers in several dis- large, dynamic real-world databases: There ciplines, including AI and databases, for a are several common problems in all kinds of decade and is still an important research fron- inductive learning algorithm: (1) When a data- tier for both machine learning and database base is very large, how can these algorithms technology.2 Although a lot of work has been speed up their learning processes? (2) When a done and some commercial data-mining pack- database is not a static repository of data—for ages are available already, existing work has not example, examples can be added, deleted, or paid enough attention to the integration of changed—the induction on the example set database and knowledge base technology with cannot be a one-time process, so how can machine-learning techniques. induction algorithms deal with the changing With the World Wide Web’s emergence as a examples? (3) When some inconsistency (for large, distributed data repository and the real- example, noise) is found in a database or a ization that online transaction databases can knowledge base just produced, how can they be analyzed for commercial gains, data mining remove it? in large databases has attracted wide interest One possible way to solve those problems is from both academia and the industry and, in incremental learning, which means dividing a the meanwhile, has also uncovered new chal- large example set into a number of subsets and lenges (Ramakrishnana and Gramam 1999). treating each subset each time. Although no Data mining has its distinctive goal from relat- existing algorithms have found a complete ed fields such as machine learning, databases, solution to these problems, many research and statistics and accordingly requires distinc- efforts have been made along this direction. tive tools. An ILDB system is one such tool to For example, AQ15, AE5, ID5R (Utgoff 1989), and implement automatic knowledge acquisition the windowing technique in ID3 can be viewed from databases. as good examples of research in incremental learning. However, how to deal with the incon- Acknowledgments sistency between new data and the data previ- I am indebted to the anonymous reviewers for their ously used is still an open question. The incon- constructive and supportive comments on earlier sistency might be caused by dynamic changes versions of this article. This article is a modified and of the data attributes. extended version of the author’s invited keynote address at the 1997 International Conference on the Generally speaking, incremental induction Practical Application of Knowledge Discovery and can take more time (but less run-time space) Data Mining, London, United Kingdom, 23 to 25 because it needs to restructure decision trees or April 1997. The original keynote address has been rules when some new examples do not fit the updated to cover the current state of the art. decision trees or rules developed so far. Stronger integration of database technolo- Notes gy and data mining. One of the criticisms 1. The discovery of quantitative laws among features- from the database community about current parameters can possibly be listed as a fourth category, data-mining research is that database technol- but this discovery is more on the statistical side than ogy can contribute more than just data prepa- AI-related data mining. ration for data mining. How to integrate data- 2. The First Knowledge Discovery and Data-Mining

66 AI MAGAZINE Articles

Workshop was held in Detroit, Michigan, Langley, P. 1987. Machine Learning and Wu, X. 1993a. LFA: A Linear Forward- in August 1989 in conjunction with the Concept Formation. Machine Learning Chaining Algorithm for AI Production Sys- 1989 International Joint Conference on 2:99–102. tems. Expert Systems: The International Artificial Intelligence. Lenat, D. B. 1983. EURISKO: A Program That Journal of Knowledge Engineering 10(4): Learns New Heuristics and Domain Con- 237–242. References cepts—The Nature of Heuristics III: Pro- Wu, X. 1993b. A Prolog-Based Representa- Aggarwal, C., and Yu, P. 1998. A New gram Design and Results. Artificial Intelli- tion for Integrating Knowledge and Data. Framework for Itemset Generation. In Pro- gence 21:61–98. Informatica: An International Journal of Com- ceedings of the ACM Symposium on Princi- Lenat, D. B. 1979. On Automated Scientific puting and Informatics 17(2): 137–144. ples of Database Systems. New York: Associ- Theory Formation: A Case Study Using the Wu, X. 1993c. Inductive Learning: Algo- ation of Computing Machinery. AM Program. In Machine Intelligence 9, eds. rithms and Frontiers. Artificial Intelligence Agrawal, R., and Srikant, R. 1995. Mining J. Hayes et al. New York: Halstead. Review 7(2): 93–108. Sequential Patterns. Paper presented at the Michalski, R.; Mozetic, I.; Hong J.; and Wu, X. 1993d. The HCV Induction Algo- Eleventh International Conference on Data Lavrac, N. 1986. The Multipurpose Incre- rithm. In Proceedings of the Twenty-First Engineering (ICDE), 3–14 March, Taipei, mental Learning System AQ15 and Its Test- ACM Computer Science Conference, Taiwan. ing Application to Three Medical Domains. 168–175. New York: Association of Com- Agrawal R., and Srikant, R. 1994. Fast Algo- In Proceedings of the Fifth National Con- puting Machinery. rithms for Mining Association Rules. Paper ference on Artificial Intelligence, 1041– Wu, Y., and Craske, N. 1997. Discovery presented at the Twentieth International 1045. Menlo Park, Calif.: American Associ- from Queries. In Proceedings of the 1997 Conference on Very Large Databases, 12–15 ation for Artificial Intelligence. IEEE Knowledge and Data Engineering September, Santiago, Chile. Mitchell, T. 1977. Version Spaces: A Candi- Exchange Workshop (KDEX-97), 120–122. Agrawal, R.; Imielinski, T.; and Swami, A. date Elimination Approach to Rule Learning. Washington, D.C.: IEEE Computer Society. 1993. Mining Associations between Sets of In Proceedings of the Fifth International Zhang, T.; Ramakrishnan, R.; and Livny, M. Items in Massive Databases. In Proceedings Joint Conference on Artificial Intelligence, 1997. BIRCH: A New Data-Clustering Algo- of the ACM SIGMOD International Confer- 305–310. Menlo Park, Calif.: International rithm and Its Applications. Data Mining ence on Management of Data, 207–216. Joint Conferences on Artificial Intelligence. and Knowledge Discovery 1(2): 141–182. New York: Association of Computing Quinlan, J. R. 1993. C4.5: Programs for Zheng, Z. 1995. Constructing Nominal X- Machinery. Machine Learning. San Francisco, Calif.: of-N Attributes. In Proceedings of the 1995 Clark P., and Niblett, T. 1989. The cn2 Morgan Kaufmann. International Joint Conference on Artificial Induction Algorithm. Machine Learning Quinlan, J. R. 1986. Induction of Decision Intelligence, 1064–1070. Menlo Park, 3:261–283. Trees. Machine Learning 1:81–106. Calif.: International Joint Conferences on Dayhoff, J. E. 1990. Neural Network Architec- Ramakrishnan, N., and Grama, A. Y. 1999. Artificial Intelligence. tures: An Introduction. New York: Van Nos- Data Mining: From Serendipity to Science. trand Reinhold. IEEE Computer 32(8): 34–37. Dougherty, J.; Kohavi, R.; and Sahami, M. Shapiro, A. D. 1987. Structured Induction in 1995. Supervised and Unsupervised Dis- Expert Systems. Wokingham, U.K.: Turing cretization of Continuous Features. In Pro- Institute Press. Xindong Wu is current- ceedings of the Twelfth International Confer- Srikant, R., and Agrawal, R. 1996. Mining ly an associate professor ence on Machine Learning, 194–202. San Quantitative Association Rules in Large in computer science at Francisco, Calif.: Morgan Kaufmann. Relational Tables. In Proceedings of the the Colorado School of Fayyad, U. M., and Irani, K. B. 1992. On the ACM-SIGMOD 1996 Conference on Man- Mines. He holds a Ph.D. Handling of Continuous-Valued Attributes agement of Data, 1–12. New York: Associa- in AI from the University in Decision Tree Generation. Machine tion of Computing Machinery. of Edinburgh. Wu is the Learning 8:87–102. Tsur, D.; Ullman, J.; Abiteboul, S.; Clifton, executive editor of Fisher, D. 1996. Iterative Optimization and C.; Motwani, R.; Nestorov, S.; and Rosen- Knowledge and Informa- Simplification of Hierarchical Clusterings. thal, A. 1998. Query Flocks: A Generaliza- tion Systems and also serves on the editorial Journal of Artificial Intelligence Research tion of Association-Rule Mining. In Pro- boards of the IEEE Transactions on Knowl- 4:147–180. ceedings of the ACM SIGMOD edge and Data Engineering and Data Mining Han, J., and Fu, Y. 1999. Multiple-Level International Conference on Management and Knowledge Discovery. He is currently Association Rules. IEEE Transactions on of Data, 1–12. New York: Association of chair of the Steering Committee of the Knowledge and Data Engineering 11(5): Computing Machinery. annual Pacific-Asia Conference on Knowl- 798–805. Utgoff, P. E. 1989. Incremental Induction of edge Discovery and Data Mining. His e- mail address is [email protected]. Han, J.; Pei, J.; and Yin, Y. 2000. Mining Fre- Decision Trees. Machine Learning 4:161– quent Patterns without Candidate Genera- 186. tion. In Proceedings of the 2000 ACM-SIG- Wnek, J., and Michalski, R. S. 1994. MOD International Conference on Hypothesis-Driven Constructive Induction Management of Data (SIGMOD ‘00). New in AQ17-HCI: A Method and Experiments. York: Association of Computing Machin- Machine Learning 14:139–168. ery. Forthcoming. Wu, X. 1999. Fuzzy Interpretation of Dis- Hong, J. 1985. AE1: An Extension Matrix cretized Intervals. IEEE Transactions on Approximate Method for the General Cover- Fuzzy Systems 7(6): 753–759. ing Problem. International Journal of Comput- Wu, X. 1995. Knowledge Acquisition from er and Information Sciences 14(6): 421–437. Databases. Norwood, N.J.: Ablex.

FALL 2000 67 Intelligent Systems for Molecular Biology he ISMB conference series provides a general forum for dissemi- Tnating the latest developments in bioinformatics. ISMB is a mul- tidisciplinary conference that brings together scientists from comput- er science, mathematics, molecular biology, and statistics. Its scope includes the development and application of advanced computational methods for biological problems.

Published by The AAAI Press 445 Burgess, Drive, Menlo Park, California 94025 http://www.aaaipress.org/ 650-328-3123 • 650-321-4457 (fax)

(AAAI members may deduct 20% from list price)

ISMB-2000—San Diego, California ISMB-96—St. Louis, Missouri Proceedings of the Eighth International Proceedings of the Fourth International Conference on Intelligent Systems for Conference on Intelligent Systems for Molecular Biology Molecular Biology Edited by , Timothy Bailey, , Edited by David J. States, Pankaj Agarwal, Michael Gribskov, Thomas Lengauer, Ilya Shindyalov, Lynn Ten Eyck, , , and Randall F. Smith and Helge Weissig ISBN 1-57735-002-2 274 pp., index, $50.00 softcover ISBN 1-57735-115-0 436 pp., index, $45.00 softcover ISMB-95—Cambridge, England ISMB-99—Heidelberg, Germany Proceedings of the Third International Proceedings of the Seventh International Conference on Intelligent Systems for Conference on Intelligent Systems for Molecular Biology Molecular Biology Edited by Christopher Rawlings, Dominic Clark, Russ Altman, Edited by Thomas Lengauer, Reinhard Schneider, Peer Bork, Douglas Lawrence Hunter, Thomas Lengauer, and Brutlag, Janice Glasgow, Hans-Werner Mewes, and Ralf Zimmer ISBN 0-929280-83-0 427 pp., index, $50.00 softcover ISBN 1-57735-083-9 324 pp., index, $45.00 softcover ISMB-94—Stanford, California ISMB-98—Montréal, Quebec, Canada Proceedings of the Sixth International Proceedings of the Second International Conference on Intelligent Systems for Conference on Intelligent Systems for Molecular Biology Molecular Biology Edited by Russ Altman, Douglas Brutlag, , Edited by Janice Glasgow, Tim Littlejohn, François Major, Richard Lathrop, and David Searls Richard Lathrop, , and Christoph Sensen ISBN 1-57735-053-7 234 pp., index, $45.00 softcover ISBN 0-929280-68-7 401 pp., index, $45.00 softcover

ISMB-97—Halkidiki, Greece ISMB-93—Bethesda, Maryland Proceedings of the Fifth International Proceedings of the First International Conference on Intelligent Systems for Conference on Intelligent Systems Molecular Biology for Molecular Biology Edited by Terry Gaasterland, Peter Karp, Kevin Karplus, Edited by Lawrence Hunter, David Searls, and Jude Shavlik Christos Ouzounis, , and ISBN 0-929280-47-4 468 pp., index, $45.00 softcover ISBN 1-57735-022-7 382 pp., index, $50.00 softcover

68 AI MAGAZINE