Attribute Reduction with Test Cost Constraint.pdfVIP

JOURNALOFELECTRONICSCIENCEAND TECHNOLOGY,VOL．9，NO．2，JUNE2011 97 AttributeReductionwithTestCostConstraint FanM inandW illiam Zhu Abstractin many machinelearning applications， builttoproducethemaximalclassificationaccuracy? dataarenotfree，andthereisatestcostforeachdata To define the problem succinctly， we use the item ．Fortheeconomicalreason，someexisting works conditionalinformationentropyoftheselectedtestsetfi_e．， try to minim ize the testcostand atthe sam etime， attributesubset)withrespecttotheclassfi．e．．decision preserve a particular property of a given decision attribute)。insteadofhteclassificationaccuracy，toevaluate system ．Inthispape~wepointoutthatthetestcostone thequalityofthetestset．W eraguehtatthisalternativeis can afr0rdislimited in someapplications．Hence．one both applicable and necessary．First，the conditional hastosacrificerespectivepropertiestokeepthetestcost informationenrtopyreflectshtedistributionofthedataset underabudget．Toofrmalizethisissue．wedefinethe over the class，and a classifier generally has better classification accuracy on a datasetwith lessentropy． testcostconstraintattributereduction problem，where Second．theclassificationaccuracyonunseenobjectsis theoptimizationobjectiveistominimizetheconditionai hardtoevaluate；itreliesonbothhtedatadistributionand information entropy．This problem is an essential theclassifier．whichhasnumerouschoicest卜【 generalization ofboth the test-cost-sensitive attribute ．