




已阅读5页,还剩37页未读, 继续免费阅读
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
IntroductiontoBoostedTrees TianqiChenOct 222014 Outline ReviewofkeyconceptsofsupervisedlearningRegressionTreeandEnsemble WhatareweLearning GradientBoosting HowdoweLearn Summary ElementsinSupervisedLearning Notations i thtrainingexampleModel howtomakepredictiongiven includelinear logisticregression canhavedifferentinterpretations Linearmodel ThepredictionscoredependingonthetaskLinearregression Logisticregression isthepredictedscoreispredictedtheprobability oftheinstancebeingpositiveOthers forexampleinrankingcanbetherankscoreParameters thethingsweneedtolearnfromdataLinearmodel Elementscontinued ObjectiveFunction Objectivefunctionthatiseverywhere Lossontrainingdata Squareloss Logisticloss Regularization howcomplicatedthemodelis L2norm L1norm lasso TrainingLossmeasureshowwellmodelfitontrainingdata Regularization measurescomplexityofmodel Puttingknownknowledgeintocontext Ridgeregression Linearmodel squareloss L2regularizationLasso Linearmodel squareloss L1regularizationLogisticregression Linearmodel logisticloss L2regularizationTheconceptualseparationbetweenmodel parameter objectivealsogivesyouengineeringbenefits ThinkofhowyoucanimplementSGDforbothridgeregressionandlogisticregression ObjectiveandBiasVarianceTrade off Whydowewanttocontaintwocomponentintheobjective OptimizingtraininglossencouragespredictivemodelsFittingwellintrainingdataatleastgetyouclosetotrainingdatawhichishopefullyclosetotheunderlyingdistributionOptimizingregularizationencouragessimplemodelsSimplermodelstendstohavesmallervarianceinfuturepredictions makingpredictionstable TrainingLossmeasureshowwellmodelfitontrainingdata Regularization measurescomplexityofmodel Outline ReviewofkeyconceptsofsupervisedlearningRegressionTreeandEnsemble WhatareweLearning GradientBoosting HowdoweLearn Summary RegressionTree CART regressiontree alsoknownasclassificationandregressiontree DecisionrulessameasindecisiontreeContainsonescoreineachleafvalue Input age gender occupation age 15 ismale 2 1 0 1 Y N Y N Doesthepersonlikecomputergames predictionscoreineachleaf RegressionTreeEnsemble age 15 ismale 2 1 0 1 Y N Y N Y N 0 9 0 9 tree1 tree2UseComputerDaily f 2 0 9 2 9f 1 0 9 1 9Predictionofissumofscorespredictedbyeachofthetree TreeEnsemblemethods Verywidelyused lookforGBM randomforest AlmosthalfofdataminingcompetitionarewonbyusingsomevariantsoftreeensemblemethodsInvarianttoscalingofinputs soyoudonotneedtodocarefulfeaturesnormalization Learnhigherorderinteractionbetweenfeatures Canbescalable andareusedinIndustry Putintocontext ModelandParameters Model assumingwehaveKtrees Think regressiontreeisafunctionthatmapstheattributestothescoreParametersIncludingstructureofeachtree andthescoreintheleafOrsimplyusefunctionasparameters Insteadlearningweightsin wearelearningfunctions trees SpaceoffunctionscontainingallRegressiontrees Learningatreeonsinglevariable Howcanwelearnfunctions Defineobjective loss regularization andoptimizeit Example Considerregressiontreeonsingleinputt time IwanttopredictwhetherIlikeromanticmusicattimet t 2011 03 01 Y N t 2010 03 20Y 0 2 Equivalently Themodelisregressiontreethatsplitsontime N1 2 1 0 Piecewisestepfunctionovertime Learningastepfunction Thingsweneedtolearn Objectiveforsinglevariableregressiontree stepfunctions TrainingLoss Howwillthefunctionfitonthepoints Regularization Howdowedefinecomplexityofthefunction Numberofsplittingpoints l2normoftheheightineachsegment SplittingPositionsTheHeightineachsegment Learningstepfunction visually Comingback ObjectiveforTreeEnsemble Model assumingwehaveKtreesObjective Possiblewaystodefine Numberofnodesinthetree depthL2normoftheleafweights detailedlater Trainingloss ComplexityoftheTrees ObjectivevsHeuristic Whenyoutalkabout decision trees itisusuallyheuristicsSplitbyinformationgainPrunethetreeMaximumdepthSmooththeleafvaluesMostheuristicsmapswelltoobjectives takingtheformal objective viewletusknowwhatwearelearningInformationgain traininglossPruning regularizationdefinedby nodesMaxdepth constraintonthefunctionspaceSmoothingleafvalues L2regularizationonleafweights RegressionTreeisnotjustforregression Regressiontreeensembledefineshowyoumakethepredictionscore itcanbeusedforClassification Regression Ranking Italldependsonhowyoudefinetheobjectivefunction Sofarwehavelearned UsingSquarelossWillresultsincommongradientboostedmachineUsingLogisticlossWillresultsinLogitBoost Outline ReviewofkeyconceptsofsupervisedlearningRegressionTreeandEnsemble WhatareweLearning GradientBoosting HowdoweLearn Summary TakeHomeMessageforthissection Bias variancetradeoffiseverywhereTheloss regularizationobjectivepatternappliesforregressiontreelearning functionlearning WewantpredictiveandsimplefunctionsThisdefineswhatwewanttolearn objective model Buthowdowelearnit Nextsection SoHowdoweLearn Objective WecannotusemethodssuchasSGD tofindf sincetheyaretrees insteadofjustnumericalvectors Solution AdditiveTraining Boosting Startfromconstantprediction addanewfunctioneachtime Modelattrainingroundt Newfunction Keepfunctionsaddedinpreviousround AdditiveTraining Considersquareloss Howdowedecidewhichftoadd Optimizetheobjective ThepredictionatroundtisThisiswhatweneedtodecideinroundt Goal find tominimizethis Thisisusuallycalledresidualfrompreviousround TaylorExpansionApproximationofLoss GoalSeemsstillcomplicatedexceptforthecaseofsquarelossTakeTaylorexpansionoftheobjectiveRecallDefine Ifyouarenotcomfortablewiththis thinkofsquarelossComparewhatwegettopreviousslide OurNewGoal Objective withconstantsremovedwhereWhyspendingsmucheffortstoderivetheobjective whynotjustgrowtrees Theoreticalbenefit knowwhatwearelearning convergenceEngineeringbenefit recalltheelementsofsupervisedlearningandcomesfromdefinitionoflossfunctionThelearningoffunctiononlydependontheobjectiveviaandThinkofhowyoucanseparatemodulesofyourcodewhenyouareaskedtoimplementboostedtreeforbothsquarelossandlogisticloss Refinethedefinitionoftree Wedefinetreebyavectorofscoresinleafs andaleafindexmappingfunctionthatmapsaninstancetoaleaf age 15 ismale Y N Y N Leaf1 Leaf2 Leaf3 q 1 q 3 w1 2 w2 0 1 w3 1 Thestructureofthetree Theleafweightofthetree DefineComplexityofaTree cont Definecomplexityas thisisnottheonlypossibledefinition Numberofleaves L2normofleafscores age 15 ismale Y N Y N Leaf1 Leaf2 Leaf3 w1 2 w2 0 1 w3 1 RevisittheObjectives DefinetheinstancesetinleafjasRegrouptheobjectivebyeachleaf ThisissumofTindependentquadraticfunctions TheStructureScore TwofactsaboutsinglevariablequadraticfunctionLetusdefine Assumethestructureoftree q x isfixed theoptimalweightineachleaf andtheresultingobjectivevalueare Thismeasureshowgoodatreestructureis TheStructureScoreCalculation age 15 ismale Y N Y N Instanceindex 1 2 3 4 5 g1 h1 g2 h2 g3 h3 g4 h4 g5 h5 gradientstatistics Thesmallerthescoreis thebetterthestructureis SearchingAlgorithmforSingleTree EnumeratethepossibletreestructuresqCalculatethestructurescorefortheq usingthescoringeq Findthebesttreestructure andusetheoptimalleafweight But therecanbeinfinitepossibletreestructures GreedyLearningoftheTree Inpractice wegrowthetreegreedilyStartfromtreewithdepth0Foreachleafnodeofthetree trytoaddasplit Thechangeof objectiveafteraddingthesplitis Remainingquestion howdowefindthebestsplit thescoreofleftchildthescoreofifwedonotsplitthescoreofrightchild Thecomplexitycostbyintroducingadditionalleaf EfficientFindingoftheBestSplit Allweneedissumofgandhineachside andcalculateLefttorightlinearscanoversortedinstanceisenoughtodecidethebestsplitalongthefeature g1 h1 g4 h4 g2 h2 g5 h5 g3 h3 Whatisthegainofasplitrule Sayisagea AnAlgorithmforSplitFinding Foreachnode enumerateoverallfeaturesForeachfeature sortedtheinstancesbyfeaturevalueUsealinearscantodecidethebestsplitalongthatfeatureTakethebestsplitsolutionalongallthefeaturesTimeComplexitygrowingatreeofdepthKItisO ndKlogn oreachlevel needO nlogn timetosortTherearedfeatures andweneedtodoitforKlevelThiscanbefurtheroptimized e g useapproximationorcachingthesortedfeatures Canscaletoverylargedataset WhataboutCategoricalVariables SometreelearningalgorithmhandlescategoricalvariableandcontinuousvariableseparatelyWecaneasilyusethescoringformulawederivedtoscoresplitbasedoncategoricalvariables Actuallyitisnotnecessarytohandlecategoricalseparately Wecanencodethecategoricalvariablesintonumericalvectorusingone hotencoding Allocatea categoricallengthvector Thevectorwillbesparseiftherearelotsofcategories thelearningalgorithmispreferredtohandlesparsedata PruningandRegularization Recallthegainofsplit itcanbenegative WhenthetraininglossreductionissmallerthanregularizationTrade offbetweensimplicityandpredictivnessPre stoppingStopsplitifthebestsplithavenegativegainButmaybeasplitcanbenefitfuturesplits Post PrunningGrowatreetomaximumdepth recursivelyprunealltheleafsplitswithnegativegain Recap BoostedTreeAlgorithm AddanewtreeineachiterationBeginningofeachiteration calculateUsethestatisticstogreedilygrowatreeAddtothemodelUsually insteadwedoiscalledstep sizeorshrinkage usuallysetaround0 1Thismeanswedonotdofulloptimizationineachstepandreservechanceforfuturerounds ithelpspreventoverfitting Outline ReviewofkeyconceptsofsupervisedlearningRegressionTreeandEnsemble WhatareweLearning GradientBoosting HowdoweLearn Summary Questionstocheckifyoureallygetit Howcanwebuildaboostedtreeclassifiertodoweightedregressionproblem suchthateachinstancehaveaimportanceweight Backtothetimeseriesproblem ifIwanttolearnstepfunctionsovertime Isthereotherwaystolearnthetimesplits otherthanthetopdownsplitapproach Questionstocheckifyoureallygetit Howcanwebuildaboostedtreeclassifiertodoweightedregressionproblem suchthateachinstancehaveaimportanceweight Defineobjective calculate feedittotheoldtreelearningalgorithmwehaveforun weightedversion Againthinkofseparationofmodelandobjective howdoesthetheorycanhelpbetterorganizingthemachinelearningtoolkit Questionstocheckifyoureallygetit Timeseriesproblem AllthatisimportantisthestructurescoreofthesplitsTop downgreedy sameastreesBottom upgreedy startfromindividualpointsaseachgroup greedilymergeneig
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年中国航天科工集团校园招聘笔试题库附答案
- 地表耕耘专业知识培训内容课件
- 地理知识简解培训内容课件
- 2025年上海短期劳动合同相关规定
- 2025年机关事务管理局机关人事处招聘笔试专项练习含答案
- 2025年高温防暑试题及答案
- 2025关于国际贸易合同的新规定
- 2025年甘肃省陇南市事业单位工勤技能考试题库(含答案)
- 2025合作加盟协议范本
- 2025年无固定期限合同解除策略详解
- 金融科技推动新质生产力发展
- 肝脓肿合并糖尿病业务查房
- 实验室安全教育考试题库实验室安全考试题库及答案
- 企业员工职业道德考核制度
- 公司安全事故隐患内部举报、报告奖励制度
- 【初中物理】质量与密度练习题 2024-2025学年初中物理人教版八年级上册
- 南外初中小语种课程设计
- 【上海市塑料探究所企业员工激励机制存在的问题及优化建议探析(论文)8200字】
- Unit2 Whats your hobby-教案人教精通版英语六年级上册
- 【必刷题】2024五年级英语上册一般过去时专项专题训练(含答案)
- T-CTSS 86-2024 原味茶饮料标准
评论
0/150
提交评论