Model order selection for boolean matrix factorization

《Model order selection for boolean matrix factorization》由会员上传分享，免费在线阅读，更多相关内容在行业资料-天天文库。

1、ModelOrderSelectionforBooleanMatrixFactorizationPauliMiettinenJillesVreekenMaxPlanckInstituteforInformaticsDept.ofMathematicsandComputerScienceSaarbrücken,GermanyUniversiteitAntwerpen,Belgiumpmiettin@mpi-inf.mpg.dejilles.vreeken@ua.ac.beABSTRACTcalledalow-dimensionalreprese

2、ntationofthedata,andisusuallyobtainedusingsomeformofmatrixfactorization.Matrixfactorizations—whereagivendatamatrixisapproximatedInmatrixfactorizationstheinputdata(representedasamatrix)byaproductoftwoormorefactormatrices—arepowerfuldataisdecomposedintotwo(ormore)factormatric

3、es.Usuallytheaimminingtools.Amongothertasks,matrixfactorizationsareoftenistohavelow-dimensionalfactormatriceswhoseproductapproxi-usedtoseparateglobalstructurefromnoise.This,however,requiresmatestheoriginalmatrixwell.Byimposingdifferentconstraints,solvingthe‘modelorderselect

4、ionproblem’ofdeterminingwhereoneobtainsdifferentfactorizations.Perhapsthetwobest-knownfine-grainedstructurestops,andnoisestarts,i.e.,whatistheproperfactorizationsareSingularValueDecomposition(SVD),closelysizeofthefactormatrices.relatedtoPrincipalComponentAnalysis(PCA),andNon

5、-negativeBooleanmatrixfactorization(BMF)—wheredata,factors,andMatrixFactorization(NMF).SVDandPCArestrictthefactorma-matrixproductareBoolean—hasreceivedincreasedattentionfromtricestobeorthogonal,whileNMFrequiresthedataandthefactorthedataminingcommunityinrecentyears.Thetechni

6、quehasmatricestobenon-negative.desirableproperties,suchashighinterpretabilityandnaturalsparsity.WhentheinputdataisBoolean,(thatis,containsonly0sand1s,ButsofarnomethodforselectingthecorrectmodelorderforBMFasistypicalwithsupermarketbasketdata),onecanapplyBooleanhasbeenavailab

7、le.InthispaperweproposetousetheMinimumMatrixFactorization(BMF).SimilarlytoNMF,itrestrictsthefactorDescriptionLength(MDL)principleforthistask.Besidessolvingmatricesforaddedinterpretabilityandsparsity.InBMF,thefactortheproblem,thiswell-foundedapproachhasnumerousbenefits,e.g.,m

8、atricesarerequiredtobeBoolean,i.e.,containonly0sand1s.itisautomatic,doesnotrequire