Title of article
A meta-heuristicapproachforimprovingtheaccuracyinsome classification algorithms
Author/Authors
Huy NguyenAnhPham، نويسنده , , EvangelosTriantaphyllou، نويسنده ,
Issue Information
ماهنامه با شماره پیاپی سال 2011
Pages
16
From page
174
To page
189
Abstract
Currentclassificationalgorithmsusuallydonottrytoachieveabalancebetweenfittingand
generalizationwhentheyinfermodelsfromtrainingdata.Furthermore,currentalgorithmsignore
the factthattheremaybedifferentpenaltycostsforthefalse-positive,false-negative,andunclassifiable
types. Thus,theirperformancemaynotbeoptimalormayevenbecoincidental.Thispaperproposesa
meta-heuristicapproach,calledtheConvexityBasedAlgorithm(CBA),toaddresstheseissues.Thenew
approachaimsatoptimallybalancingthedatafittingandgeneralizationbehaviorsofmodelswhen
sometraditionalclassificationapproachesareused.TheCBAfirstdefinesthetotalmisclassificationcost
(TC) asaweightedfunctionofthethreepenaltycostsandthecorrespondingerrorratesasmentioned
above. Nextitpartitionsthetrainingdataintoregions.Thisisdoneaccordingtosomeconvexity
propertiesderivablefromthetrainingdataandthetraditionalclassificationmethodtobeusedin
conjunctionwiththeCBA.NexttheCBAusesageneticapproachtodeterminetheoptimallevelsof
fitting andgeneralization.The TC is usedasthefitnessfunctioninthisgeneticapproach.Twelvereal-
life datasetsfromawidespectrumofdomainswereusedtobetterunderstandtheeffectivenessofthe
proposedapproach.ThecomputationalresultsindicatethattheCBAmaypotentiallyfillinacriticalgap
in theuseofcurrentorfutureclassificationalgorithms.
Keywords
classification , Fitting , False positive , Generalization , False negative , Unclassifiable , Convex region , Genetic algorithms , Optimization
Journal title
Computers and Operations Research
Serial Year
2011
Journal title
Computers and Operations Research
Record number
927845
Link To Document