Weintroducecoactive learning as a distributed learning approachto data miningin networkedand distributed databases. Thecoactive learningalgorithmsact on independent data sets and cooperatebycommunicatingtraining information, whichis usedto guidethe algorithms'hypothesisconstruction. Theexchangedtraining informationis limited to examplesandresponsesto examples.It is shownthat coactive learningcanoffer a solution to learningonverylarge data sets byallowingmultiplecoactingalgorithmsto learnin parallel onsubsetsof the data,evenif thesubsetsare distributed overa network.Coactivelearningsupportsthe construction of global concept descriptions evenwhenthe individual learning algorithmsare providedwithtraining sets having biasedclass distributions.Finally,the capabilitiesof coactive learningare demonstratedonartificial noisydomains,andon real worlddomaindata withsparseclass representationand unknownattribute values.
Dan L. Grecu, Lee A. Becker