Addressing Knowledge Discovery Problems in a Multistrategy Framework

ion with equal frequency. Cognitive scientists speak of bask level nodes within a generalization hierarchy whose children share many sensorially recognizable commonalities (Rosch et al. 1976). Other factors that help to characterize a node’s utility compared to those at higher or lower levels of abstraction are concept typicality (bow common are the features of this concept among its sibling concepts), and the coutext in which the concept is being used (gnmesch 1988; Kubat, Bratko & Michalski 1996). Each of these factors affects the selection of a particular level of abstraction in m~i~ clescriptkms. By encoding the relative utility of the nodes into the knowledge representation of a discovery system, the system can wesent discovered knowledge that focuses on the mm~ useful levels of abstraction wheat possible. We will typically prefer to see the classification roles "An From: Proceedings of the Third International Conference on Multistrategy Learning. Copyright © 1996, AAAI (www.aaai.org). All rights reserved. object belongs to Class 1 if it is an apple" or "An object belongs to Class 1 if it is a fruit" instead of "An object belongs to Class 1 if it is a red delicious" of "An object belongs to Class I if it is food." Another important feature in INLEN’s tructured dam representation is its ability to wcfk with multiple views of the data (Km~frmm & Mi¢lmlski, 1996b). Consider application in which a marketing specialist is trying to target the cnst~n~rs who are most likely, to be inlerested in a new product. A oustomerdatabase may have extensive inf~ou including the model of the automobile driven by the customer, Au~obiies may in turn be Ol’~aniT~l according to manufacun~,:type (e,g., sedan, sporm car, statiou wagon), ltmce, year, etc. One may not know prior to a kamin8 taskwhich classificatiou view will provide the most concise and useful knowledge. Also, more than one view may generate the best rules for detenni~g whether a customer is. likely to buy the product. A decision rule, for example, may include the condition that likely buyer will often drive European station wagons, while the specificatiou of the type of vehicle or manufacturer alone may not give an accurate representation of the customer’s I~Opel~ties. I~.l~q’S gene~liTm_ion engine aotommie~lly selects from the possible ways of expressing a set of attn~te values (in.this case the automobile medeis.that are European.station wagons)a concise representatiou of the knowledge. INLEN’s knowledge generation operators have also been e~hRnced to cope with the problem of incomplete data sets. Expe~’iments with sparse ~ exposed some of the limi~ion$ of these opelators, leading to modificgfious in which the logical implications of unknowns are more rigoronsly encoded into the learning algoritlnn. For example, in the AQ fmnily of programs, attributes with unknown values are represented as .having all of their values under consideration. A stipulation in earlier versionsof the .program :required that gene~li,Atious of examples with unknown values for some attributes maintain consistency by permitting those amamtes to lake any value. While this wil! guarantee rigorous consistency with the data, trouble arises when many examples have only a few known attribute values. Gene~liztng just a few of them together creates a siVmfion in which noshing can he assumed about Any featore. As an e~le of a donmin in which this may be a real problem, intelligent agcuts, me being.developed to scan text and smnmafize it based on key words in several categories of intea~st to the user. Articles often will oaly COBtMn key words in a few of these.categedes, leading to a very empty database. The relationships amoug entries in the various categeries will often be.tenuoas ones given.that for much of the data, oue.or mine of these fields will.be empty. In such a domain a discovery program must be able to sift through the information that is present without getting lost in that which is missing. Atthe expense of some additional computational complexity, the knowledge generation operators in I~-~:N have been modified in such a way thin they can now generate knowledge consistent with what facts have been made available, without ~dhering to the assmnption tlwt unknowns must be gcueralized to take on all values. With the relaxation of this condition, a learning engine.osndete~ Inore Szlbstantial relati0o~hins. Another aspect of this research approw,~s the.problem ̄ of knowledgeextraction from distributed sources. The INSIGHT program (Ribeim, Kaufma~ & Kerschberg. 1995) is being developed as aa operator to perform knowledg~dtiven search through multiple databases, The combinatorial cost of combining separate data sets is avoided through INSIGHT’s mechanism of finding relationships .between a database and the knowledge generated frem mmther database, In order to facilitate the interface with INSIGHT and Other opera~rs, ]NLEN mRinmins informatiou on the database reco~ relevant to each rule in its knowledge base. As was shown above in the population growth example, a.use for.these linh i8 tO enabie Ihe idcutifm:atiou of significant cinstets of exceptious. Another use is to allow the measurement of the degree of match between a role and a set of recot~ in a second database mmehing a given set of mnditious. A high degree of retch my suggest a ]ink~e between the two co~-pts. For example, a role describing the clima~ of a country may be crossreferenced with a database of natural di~stets. Ifa class of nateraldisaster occurred in a set of countries gimilar to I~ set of countries covered by the cllm~ rule, it may suggest a relatioaship betwem the clim~ and that kind of disaster.