




版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
FeatureSelectionforRanking
Cenfangming2007.10ABSTRACT
TherealityisthatmanyfeatureselectionmethodsusedinclassificationaredirectlyappliedtorankingWearguethatbecauseofthestrikingdifferencesbetweenrankingandclassification,itisbettertodevelopdifferentfeatureselectionmethodsforranking.
InthispaperweproposeanewfeatureselectionmethodDefinetherankingaccuracyintermsofaperformancemeasureoralossfunctionastheimportanceofthefeatureandthecorrelationbetweentherankingresultsoftwofeaturesasthesimilaritybetweenthem.Wealsodemonstratehowtosolvetheoptimizationprobleminanefficientway.Wehavetestedtheeffectivenessofourfeatureselectionmethodontwoinformationretrievaldatasetsandwithtworankingmodel1.INTRODUCTION
RankingSVMandRankNet
First,featureselectioncanhelpenhanceaccuracyinmanymachinelearningproblems,whichstronglyindicatesthatfeatureselectionisalsonecessaryforranking
Second,featureselectioncanalsohelpimprovetheefficiencyoftraining.Ininformationretrieval,especiallyinwebsearch,usuallythedatasizeisverylargeandthustrainingofrankingmodelsiscomputationallycostly.
Therehavebeennomethodsoffeatureselectiondedicatedlyproposedforranking.Mostofthemethodsusedinrankingweredevelopedforclassification.Featureselectionmethodsinclassificationfallintothreecategorie
Thefirstcategory,whichisnamedfilter,featureselectionisdefinedasapreprocessingstepandcanbeindependentfromlearning.thatinformationgain(IG)andchi-square(CHI)areamongthemosteffectivemethodsoffeatureselectionforclassification.
ThesecondcategoryreferredtoaswrapperutilizesthelearningsystemasablackboxtoscoresubsetsoffeaturesThethirdcategorycalledtheembeddedmethodperformsfeatureselectionwithintheprocessoftraining.Second,theevaluationmeasures(e.g.meanaverageprecision(MAP)andnormalizeddiscountedcumulativegain(NDCG))usedinrankingproblemsaredifferentfromthosemeasuresusedinclassification
1)inrankingusuallyprecisionismoreimportantthanrecallwhileinclassificationbothprecisionandrecallareimportant;
SeveralproblemsmayarisewhenapplyingthefeatureselectionmethodstorankingFirst:existingfeatureselectionmethodsforclassificationarenotsuitableforranking.Inranking,anumberoforderedcategoriesareused,representingtherankingrelationshipbetweeninstances,whileinclassificationthecategoriesare“flat”
2)
inrankingcorrectlyrankingthetop-ninstancesismorecriticalwhileinclassificationmakingacorrectclassificationdecisionisofequalsignificanceforallinstances.
Thefollowingarethepropertiesofthenovelmethodforfeatureselectioninranking1)Themethodmakesuseofrankinginformation,insteadofsimplyviewingtheranksasflatcategories.2)Inspiredbytheworkin[1][14][27],itconsidersthesimilaritiesbetweenfeatures,andtriestoavoidselectingredundantfeatures.3)Itmodelsfeatureselectionforrankingasamulti-objectiveoptimizationproblem.Thefinalobjectiveistofindasetoffeatureswithmaximumimportanceandminimumsimilarity.4)Itprovidesagreedysearchalgorithmtosolvetheoptimizationproblem.Thecorrespondingsolutionproducedisproventobeequivalenttotheoptimalsolutiontotheoriginalproblemundercertaincondition.2.FEATURESELECTIONMETHOD
2.1OverviewSupposethegoalistoselect(1≤t≤m)featuresfromtheentirefeatureset{1,2,…,}.Inourmethodwefirstdefinetheimportancescoreofeachfeature,anddefinethesimilaritybetweenanytwofeaturesand.Thenweemployanefficientalgorithmtomaximizethetotalimportancescoresandminimizethetotalsimilarityscoresofasetoffeatures.
2.2ImportanceoffeatureWefirstassignanimportancescoretoeachfeature.Specifically,weproposeusinganevaluationmeasurelikeMAPandNDCGoralossfunction(e.g.pair-wiserankingerrors[10][13])tocomputetheimportancescore.Intheformer,wefirstrankinstancesusingthefeature,evaluatetheperformanceintermsofthemeasure,andthentaketheevaluationresultastheimportancescore.Inthelatter,wealsorankinstancesusingthefeature,andthenviewascoreinverselyproportionaltothecorrespondinglossastheimportancescore
2.3SimilaritybetweenfeaturesInthiswork,wemeasurethesimilaritybetweenanytwofeaturesonthebasisoftheirrankingresults.Thatis,weregardeachfeatureasarankingmodel,andthesimilaritybetweentwofeaturesisrepresentedbythesimilaritybetweentherankingresultsthattheyproduce.wechooseKendall’s𝜏asanexample.TheKendall’s𝜏valueofqueryqforanytwofeatures𝑣𝑖and𝑣𝑗canbecalculatedasfollows,
2.4Optimizationformulation
wewanttoselectthosefeatureswithlargesttotalimportancescoresandsmallesttotalsimilarityscores.Mathematically,thiscanberepresentedasfollows:
In(1),therearetwoobjectives:tomaximizethesumoftheimportancescoresofindividualfeatures,andtominimizethesumofsimilarityscoresbetweenanytwofeatures.wetakeacommonapproachinoptimizationandconvertmulti-objectiveprogrammingtosingleobjectiveprogrammingusinglinearcombination.
2.5Solutiontooptimizationproblem
Theoptimizationin(2)isatypical0-1integerprogrammingproblem.Onepossibleapproachwouldbetoperformexhaustivesearch.However,thetimecomplexityofit,𝑂(𝐶𝑚),istoohightomakeitapplicableinrealapplications.Weneedtolookformorepracticalsolutions.
Inthiswork,weproposeagreedysearchalgorithmfortacklingtheissueAlgorithmGAS(GreedysearchAlgorithmoffeatureSelection)1.ConstructanundirectedgraphG0,inwhicheachnoderepresentsafeature,theweightofnode𝑣is𝜔andtheweightofanedgebetweennode𝑣andnode𝑣is 2.ConstructasetStocontaintheselectedfeatures.InitiallyS0=∅.
3.Fori=1…t,(1)Selectthenodewiththelargestweight,withoutlossofgenerality,supposethattheselectednodeis
(2)Apunishmentisconductedonalltheothernodesaccordingtotheirsimilaritieswith𝑣.Thatis,theweightsofalltheothernodesareupdatedasfollows.(3)Add𝑣tothesetSandremoveitfromgraphGtogetherwithalltheedgesconnectedtoit:4.OutputSt.
Theorem1:WiththegreedysearchalgorithminFig.1onecanfindtheoptimalsolutiontoproblem(2),providedthat𝑆𝑡𝑡,where𝑆denotestheselectedfeaturesetwith𝑡Proof:Thecondition𝑆𝑡indicatesthatwhenselectingthe(t+1)-thfeature,wedonotchangethealready-selectedtfeatures.Denote𝑆𝑡={𝑣𝑘|𝑖i=1,…,t},where𝑣istheki-thfeatureselectedinthei-thiteration.Thenthetaskturnsouttobethatoffindingthe(t+1)-thfeaturesothatthefollowingobjectivecanbemet.
Since,wecanrewrite(3)as
Andsince𝑆𝑡and={𝑣|i=1,…,t𝑡},(4)equals 3.EXPERIMENTSETTINGS
3.1DatasetsInourexperiments,weusedtwobenchmarkdatasets.1)Thefirstdatasetisthe.govdatawhichwasusedinthetopicdistillationtaskofWebtrackofTREC2004[28].2)TheseconddatasetistheOHSUMEDdata[9],whichwasusedinmanyexperimentsininformationretrieval[6][10],includingtheTREC-9filteringtrack[26].3.2Evaluationmeasures
3.2.1Meanaverageprecision(MAP)
MAPisameasureonprecisionofrankingresultsItisassumedthattherearetwotypesofdocuments:positiveandnegative(relevantandirrelevant).Precisionatnmeasurestheaccuracyoftopnresultsforaquery.Averageprecisionofaqueryiscalculatedbasedonprecisionatn:
3.2.2Normalizeddiscountcumulativegain(NDCG)
NDCGisdesignedformeasuringrankingaccuracieswhentherearemultiplelevelsofrelevancejudgment.Givenaquery,NDCGatpositionninisdefined3.3Rankingmodel3.3.1RankingSVM
RankingSVMmakesanextensionofSVMtoranking;incontrasttotraditionalSVMwhichworksoninstances,RankingSVMutilizesinstancepairsandtheirpreferencelabelsintraining.TheoptimizationformulationofRankingSVMisasfollows:
3.3.2RankNet
RankNetalsousesinstancepairsintraining.itemploysaneuralnetworkastherankingfunctionandrelativeentropyaslossfunction.Letbetheestimatedposteriorprobability𝑃andbethe“true”posteriorprobability,andlet.ThelossforaninstancepairinRankNetisdefinedas
3.4Algorithmsforcomparison
Ourproposedalgorithmhastwovariants.Welisttheminthefollowingtable.
Forcomparison,weselectedIGandCHIasthebaselines.IGmeasuresthereductioninuncertainty(entropy)inclassificationpredictionwhenknowingthefeature.CHImeasuresthedegreeofindependencebetweenthefeatureandthecategories.wealsoused“WithAllFeatures(WAF)”asanotherbaseline,inordertoshowthebenefitofconductingfeatureselection.4.EXPERIMENTALRESULTS
4.1The.govdataFig.2showstheperformancesofthefeatureselectionmethodsonthe.govdatasetwhentheyworkaspreprocessorsofRankingSVM.Fig.3showstheperformanceswhenusingRankNetastherankingmodel.Inthefigures,thex-axisrepresentsthenumberofselectedfeatures.ExperimentalresultsindicatethatinmostcasesGAS-LcanoutperformGAS-E,althoughnotsignificantlyExperimentalresultsalsoindicatethatwithGAS-LandGAS-EasfeatureselectionmethodstherankingperformancesofRankingSVMaremorestablethanthosewithIGandCHIasfeatureselectionmethods.4.2OHSUMEDdataFig.4showstheresultsofdifferentfeatureselectionmethodsontheOHSUMEDdatasetwhentheyworkaspreprocessorsofRankingSVMFig.5showstheresultsofdifferentfeatureselectionmethodsontheOHSUMEDdatasetwhentheyworkaspreprocessorsofRankNet4.3Discussions
Fromtheresultsofthetwodatasets,wemadethefollowingobservations:1)Featureselectioncanimprovetherankingperformancemoresignificantlyforthe.govdatasetthanfortheOHSUMEDdataset.
2)OurproposedalgorithmsoutperformIGandCHImoresignificantlyforthe.govdatasetthanfortheOHSUMEDdataset.
Tofigureoutthereasons,weconductedthefollowingadditionalexperiments.WefirstplottedtheimportanceofeachfeatureinthetwodatasetsinFig.6.Thex-axisrepresentsfeaturesandthey-axisrepresentstheirMAPvalueswhentheyareregardedasrankingmodels.Furthermore,weplottedthesimilaritybetweenanytwofeatures(intermsofKendall’s𝜏)inthetwo\datasetsinFig.7.Here,bothx-axisandy-axisrepresentfeatures,andthelevelofdarknessrepresentsthestrengthofsimilarity(thedarker,themoresimilar).Basedonthediscussionsabove,weconcludethatiftheeffectsoffeaturesvarylargelyandthereareredundantfeatures,ourmethodcanworkverywell.Whenapplyingourmethodinpractice,therefore,onecanfirsttestthetwoaspects.5.CONCLUSIONSANDFUTUREWORK
Inthispaper,wehaveproposedanoptimizationmethodforfeatureselectioninranking.hecontributionsofthispaperincludethefollowingpoints.1)Wehavediscussedthedifferencesbetweenclassificationandranking,andmadeclearthelimitationsoftheexistingfeatureselectionmethodswhenappliedtoranking.2)Wehaveproposedanovelmethodtoselectfeaturesforranking,inwhichtheproblemisformalizedasanoptimizationissue.3)Wehaveevaluatedtheproposedmethod
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 2025铝合金门窗销售合同
- 2025玛雅物业管理合同
- 谷物种植气候智能农业技术考核试卷
- 审议公司相关管理制度
- 审计行政人员管理制度
- 员工个人印章管理制度
- 垃圾转运现场管理制度
- 电视机防水防尘技术与应用考核试卷
- 大学浴室安全管理制度
- 埃及取消闭环管理制度
- 规范网络设备管理制度
- 2025年铁路列车员(中级)职业技能鉴定参考试题库-下(判断题)
- 电商运营岗位技能测试卷
- 2025工程建设项目多测合一成果报告书范本
- 麻醉科麻精药品PDCA管理
- 语言习得神经机制探究-深度研究
- 儿童发展问题的咨询与辅导-案例1-5-国开-参考资料
- 2025年河北石家庄市市属国有企业招聘笔试参考题库含答案解析
- 2025年国航股份地面服务部校园招聘笔试参考题库含答案解析
- 宣传物料制作合同范本
- 2025年度安徽白帝集团限公司社会招聘高频重点提升(共500题)附带答案详解
评论
0/150
提交评论