Bilenkoaddedthatthereare“noplans”tocommercialiseCatBoostorcloseitoffinanyotherproprietaryway.“It’snotaquestionofcompetitors,”hesaid.“We’dbegladtohavecompetitorsuseitasit’sfoundational.”Ofcourse,asYandexcontinuestogrow,ithaslongbeenlookingatwaysofraisingitsinternationalprofileoutsideoftheRussian-speakingworld.Moveslikethisunderscorenotjustthecompany’scommitmenttotheopensourcecommunity,butalsoit’shopetobeatthecenterofhowitdevelops,bothamonglargetechcompaniesandthelargerdevelopercommunityieltsfee.JustasGooglehascontinuedtoexpandandupdateTensorflow,theideaisthattoday’sCatBoostreleaseisthefirstiterationthatwillbeupdatedanddevelopedfurther,Bilenkotoldme.Today,thelibraryhasthreemainfeatures:“Reducedoverfitting”whichYandexsayshelpsyougetbetterresultsinatrainingprogram.Itis“basedonaproprietaryalgorithmforconstructingmodelsthatdiffersfromthestandardgradient-boostingscheme.”“Categoricalfeaturessupport”inwhichyourtrainingresultsareimprovedwhilelettingyouuseofnon-numericfactors,“insteadofhavingtopre-processyourdataorspendtimeandeffortturningittonumbers.”ItalsousesanAPIinterfacethatletsyouuseCatBoostfromthecommandlineorviaAPIforPythonorR,includingtoolsforformulaanalysisandtrainingvisualisation.Whilethereareanumberofotherlibrariesouttheretohelpwithgradientboostingorothersolutionstohelptrainmachinelearningsystems(XGBoostbeingone),BilenkoarguedthatthebenefitofCatBoostandotherframeworksputouttherebylargecompanieslikeYandexisthattheyare“battletested”foraccuracy.“Thedirtysecretwithalotofmachinelearningcodeisthatitrequiresprettyextensivetuning,”hesaid.“Oursrequireslittleandprovidesprettygoodperformanceoutofthebox.Thatisakeydifferentiator.” ManagedServices;
