classifier_cascade_for_minimizing_feature_evaluation_cost

《classifier_cascade_for_minimizing_feature_evaluation_cost》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、ClassifierCascadeforMinimizingFeatureEvaluationCostMinminChen1Zhixiang(Eddie)Xu1KilianQ.Weinberger1OlivierChapelle2DorKedem112WashingtonUniversityinSaintLouisYahoo!ResearchSaintLouis,MOSantaClara,CAchenm,zhixiang.xu,kilian,kedem.dor@wustl.educhap@yaho

2、o-inc.comAbstractareusedhundredsofmillionsoftimesperdayandarere-lieduponbybillionsofpeoplearoundtheworld.MachinelearningalgorithmsareincreasinglyHowever,thereisadistinctdifferencebetweenthemachineusedinlarge-scaleindustrialsettings.Here,thelearningsc

3、enariosintypicalresearchpapersandthereal-operationalcostduringtest-timehastobetakenworldindustrialsetting.Inindustrialsettings,theaver-intoaccountwhenanalgorithmisdesigned.Thisagecomputationalcostduringtest-timeisaseriouscon-operationalcostisaffected

4、bytheaveragerun-siderationwhenalgorithmsaredeployed.Ifataskisper-ningtimeandthecomputationtimerequiredforformedmillionsoftimesperday,itiscrucialthattheaver-featureextraction.Whenadiversesetoffeaturesagecomputationtimerequiredperinstanceissufficientlyi

5、sused,thelattercanvarydrastically.Inthislowtostaywithinthelimitsoftheavailablecomputationalpaperweproposeanalgorithmthatconstructsaresources.Asanexampleconsideracommerciale-mailcascadeofclassifierswhichexplicitlytrades-offspamfilter.Ithastoprocessmilli

6、onsofmessagesperday,operationalcostandclassifieraccuracywhileac-andgivenitslimitedresourcesmustspendlessthan10countingforon-demandfeatureextractioncosts.millisecondsoneachindividuale-mail.Similarly,awebDifferentfrompreviouswork,ouralgorithmre-searchen

7、ginemighthavetoscorehundredsofthousandsoptimizestrainedclassifiersandallowsexpen-ofdocumentswithinafewmilliseconds.sivefeaturestobescheduledatanystagewithinThetwokeydifferencesfromtraditionalmachinelearningthecascadetominimizeoverallcost.Exper-arethat

8、1.thecomputationalcostisevaluatedonaverageimentsonactualweb-searchrankingdatasetspertestinstanceand2.featuresarecomputedon-demanddemonstratethatourframeworkleadstodrasticandvarysignificantlyincost.Forexample,inthee-mailtest-timeimprovements.spamfilteri