data+science+in+talkingdata-

《data+science+in+talkingdata-》由会员上传分享，免费在线阅读，更多相关内容在行业资料-天天文库。

1、DataScienceinTalkingData主讲人：TalkingData首席数据科学家张夏天DatainTalkingDataCHINA’SLARGESTINDEPENDENTMOBILEDATAPLATFORMEstablishedin2011HeadquartersinBeijingThreeroundsofVCfinancing650mln+100,000+30mln200mln+MonthlyActiveAppswithSDKDailyMobileAdMonthlyDeviceUniqueDevicesIntegratedClicks:China

2、’sPanelonAppInstallLargestMobileAd&UsageTrackingPlatformChallengesinTalkingDataBigDataVariousApplications•Volume•Finance•Velocity•Retail•Variety•RealEstate•Variability•…•Veracity•UnreadableDataDataScienceinTalkingDataLearningonBigDataImproveEfficiencyofDataScience•Fregata•SmartDataL

3、ab•Myna•AutoModel•EventDataMiningApplicationsOpen•Lookalike•BusinessPartners•RecommenderSystem•AcademicPartners•DemographicCognition•Education•ChurnAlert•……•ContextAwareness•IndoorPositioning•……LearningonBigDataFregata(OpenSource)•LargescalemachinelearninglibraryonSparkMyna(OpenSour

4、ce)•TheframeworkofcontextawarenessofAndriodEventDataMining•Eventdatamanagementsolution•Eventdata&unreadabledataminingTheRoadToHighPerformanceMLAlgorithms:Fregata‘sApproachRemoveHypeParameters•GreedystepaveragingoptimizationmethodLowCostParallelizationMethod•Modelaveragingmethod•Conv

5、ergencewithonlyonescanofthewholedataCompressModelSizes•Expandthemodelcapabilityonasinglenodebyafactorof1000GreedyStepAveraginghttps://arxiv.org/abs/1611.03608ConvergenceofGSAGSAvsSGDGSAvsAdadeltaGSAvsSCSGParallelizationGradientAveraging/ηHighcostontrainingstage�"=�"$%−)��,(�"$%)�,01

6、ModelAveraging/1�"=)�"$%,,�,01SuitableforSparkScoreAveraging81Highcostonscoringstage�5=)�5,7�701ConvergenceofModelAveragingThemodelaveragingmethodcanapproachtheoptimalmodelforlinearproblemswithaverylargeamountoftrainingdata.Fregatavs.MLLib:LogisticRegressionFregatavs.MLLib:Softmaxon

7、MNISTModelCompressionDiscretizeparametervaluesbyK-Means•Typically,discretizeparametervaluesto128buckets.•Thenwecanuse7bitstoencodeabucket,andbuildamappingindextodiscretizeparametervalues.CompresstheresultingmodelbitmapbyRoaringBitmapsModelCompression:AccuracyCompressedModelOriginalM

8、odel(128buckets)Dat