陈天奇xgb论文。
Treeboostingisahighlyeectiveandwidelyusedmachinelearningmethod.Inthispaper,wedescribeascalableendto-endtreeboostingsystemcalledXGBoost,whichisusedwidelybydatascientiststoachievestate-of-the-artresultsonmanymachinelearningchallenges.Weproposeanovelsparsity-awarealgorithmforsparsedataandweightedquantilesketchforapproximatetreelearning.Moreimportantly,weprovideinsightsoncacheaccesspatterns,datacompressionandshardingtobuildascalabletreeboostingsystem.Bycombiningtheseinsights,XGBoostscalesbeyondbillionsofexamplesusingfarfewerresourcesthanexistingsystems.