贪心科技|让每个人享受个性化教育服务Review:PaperReadingXGBoost:AScalableTreeBoostingSystem范老师2020/04/05贪心科技|让每个人享受个性化教育服务ContentBackgroundRelatedWorkProblemDefinitionOptimizationSystemDesignExperimentApplicationConclusionReference贪心科技|让每个人享受个性化教育服务BackgroundDecisiontreeBaggingvs.boostingGBDT贪心科技|让每个人享受个性化教育服务RelatedWorkRegularizedmodelbysimplifiestheobjectiveandalgorithmforparallelizationSparsitydataprocessingOut-of-corecomputationandcache-awarepre-fetchingWeightedquantilesketch贪心科技|让每个人享受个性化教育服务ProblemDefinitionTrainedasAtthetthiterationTaylorexpressionObjective(byremovingconstant)Additivetraining叠加式训练贪心科技|让每个人享受个性化教育服务ProblemDefinition贪心科技|让每个人享受个性化教育服务ProblemDefinition贪心科技|让每个人享受个性化教育服务Optimization–Splitfinding贪心科技|让每个人享受个性化教育服务Optimization–Splitfinding3.1Exactgreedyalgorithmisrequiredtofindallpossiblesplitsonallthefeatures.[1]3.2Whendatacannotbeputintomemoryentirely,anapproximatealgorithmisneededtoreplaceexactgreedyalgorithm.[1]3.3Percentilesofafeatureisanimportantstepfortheapproximatealgorithmistoproposecandidatesplitpoints.Theauthorproposesanoveldistributedweightedquantilesketchalgorithm.[1]3.4Sparsity-awareSplitFinding:defaultdirection贪心科技|让每个人享受个性化教育服务SystemDesign1.Blockstructure2.Cache-awareaccessi.Internalbufferineachthread,toaccumulategradientstatisticsinamini-batchmanner.[1]ii.Blocksize:216examples3.Out-of-corecomputationBlockcompressionandblocksharding贪心科技|让每个人享受个性化教育服务SystemDesign贪心科技|让每个人享受个性化教育服务Experiment1.Dataset2.Classification3.Learningtorank4.Out-of-coreexperiment5.DistributedExperiment贪心科技|让每个人享受个性化教育服务Application1.Classification(e.g.highenergyphysicseventclassification;webtextclassification;malwareclassification;productcategorization;Motiondetection)二分类,计算出值之后再加一个sigmoid函数;如果是多分...