2024阿里技術(shù)參考圖冊算法篇

（（本地生活（&金融服務(wù)（螞蟻金服（菜鳥、市場營銷（阿里媽媽、云計算（阿里云）等。設(shè)施如何支撐商業(yè)系統(tǒng)？集群調(diào)度系統(tǒng)高效怎樣提升資源效率？新一代分布式臺打造的世界級計算能力、達(dá)摩院機器智能實驗室引領(lǐng)的技術(shù)和產(chǎn)業(yè)創(chuàng)新、新零售開啟的TAC/C++/如果你希望了解業(yè)界最新技術(shù)趨勢，來自阿里人工智能實驗室、天貓、淘寶、

300篇+阿里技術(shù)精華干貨（持續(xù)更新中。 人工智能實驗室研究 永阿里媽 靖達(dá)摩院機器智能實驗 智阿里云資深專 蔡商業(yè)機器智能部資深算 永營銷平臺資深算法專 志

資深算法專 三資深技術(shù)專 粵螞蟻金服資深技術(shù)專 南菜鳥網(wǎng)絡(luò)人工智能部資深算法專 元國際技術(shù)事業(yè)部資深算法專 守 智慧物流4.545.07機器學(xué)習(xí)、語音、NLP機器學(xué)習(xí)、語音、NLP

10%-CPU平均利用率業(yè)內(nèi)低于10%-CPU平均利用率業(yè)內(nèi)低于

完全兼容OSSAPI

X-X-1/10MySQLGPU計算加速，10國內(nèi)首個CloudNative單實例百萬

別率>95%,自愈時延<3秒99%/99%/99%的數(shù)據(jù)存儲與95%的計算任務(wù)60K+/60000+服務(wù)器，10作業(yè)量400阿里聯(lián)合英特爾發(fā)布的BigBenchOnMaxCompute+PAIMaxCompute2.0+$589.91$354.7 天貓精 拍立 智能地 互聯(lián)網(wǎng)汽 智能客 2017年10月AliOSThings開源，

北 杭 深 新加 以色 西雅 硅 俄羅UCBerkeley、Stanford、NTU

訪問學(xué)者計劃創(chuàng)新研究計劃

阿里巴巴技術(shù)論壇人工智能實驗室研究 永gyState-of-the-Big

Deep

Goog1eScho1ar??Y??ECVPR2A18yQ?@9? y— VOICE,

AIé[y—[Y?ó??VA??,§?cY?YVA—Ski11sof

?ù??????—?$JJ???Tma11Genie-A1ibaba′s JackMaandHongkongchiefexecutiveLin JackMaandPresidentofCommissionofChina

Ministryofscienceandtechno1ogyOf

ChineseAcademyof

????x11kī—

o.1in in rket?????

è?o?—??-A1iGenie

3D

SensorFusionFast

HardwareHome-

Car-

Persona1-??AI,阿里媽 靖

UniDesk??????1?ǚ???ǚ?-

???ǚ?-OBF;m;L76X????"?ù?áT??"?ù?áY??:Kernel??,Tree-1. 2.3.KunGai,XiaoqiangZhu,HanLi,KaiLiu,ZheWang.LearningPiece-wiseLinearModelsfromLargeScaleDataforAdClickPrediction./abs/17A4.A5194ň?????nà??DeepInterest:^?fi??ā—????:?,?—ār??DIN(2A16-ār??????ê??:únà,?"—ò?????^1?ě,???_?n(?p—

GuoruiZhou,ChengruSong,XiaoqiangZhu,XiaoMa,YanghuiYan,XingyaDai,HanZhu,JunqiJin,HanLi,KunGai.DeepInterestNetworkforClick-ThroughRatePrediction./abs/17A6.A6978????1?id—^id?ā—^???ánà??+??????oAdvancedModelServerPoweredXDL:?$??í??—AlimamaVision(OCR??2A16?6???ICAR?g?[$??āǚ

UserBehaviorSequence TargetAdTiezhengGe,LiqinZhao,GuoruiZhou,KeyuChen,ShuyingLiu,HuimingYi,ZelinHu,BochaoLiu,PengSun,HaoyuLiu,PengtaoYi,SuiHuang,ZhiqiangZhang,XiaoqiangZhu,YuZhang,KunGai.ImageMatters:VisuallymodelinguserbehaviorsusingAdvancedModelServer./abs/1711.A65A5??y— CFf=q:???á+±????ē

f三???u:??1AK? Recall ???Recall Youtube??DNN HanZhu,PengyeZhang,GuozhengLi,JieHe,HanLi,KunGai.LearningTree-basedDeepModelforRecommenderSystems./abs/18A1.A2294Optimized??:???}(a)?′Hē?,(b)???gmv,(c)??:???′H?ǒu????s,eˉě???rpmμgmv--???xs?,ecpm?èù?°?êYˉ??--?u?????--y—q?′Hμ?-u???:??μ???fHanZhu,JunqiJin,ChangTan,FeiPan,YifanZeng,HanLi,KunGai.OptimizedCostperClickinTaobaoDisplayAdvertising.InKDDJunqiJin,ChengruSong,HanLi,KunGai,JunWang,WeinanZhang.Real-TimeBiddingwithMulti-AgentReinforcementLearninginDisplayAdvertising.DiWu,XiujunChen,XunYang,HaoWang,QingTan,XiaoxunZhang,KunGai.BudgetConstrainedBiddingbyModel-freeReinforcementLearninginDisplayAdvertising.達(dá)摩院機器智能實驗 智iDSTInstituteofDataScienceandechnologiesMachineIntelligenceAI=Sensing+Learning+SensingSpeech/voiceImage/videoNaturallanguageprocessingMachineDeepReinforcementOptimization&DecisionPredictiveinventoryDeliveryassignmentManufacturingPredictiveSpeechechnologiesOurWorks–SupportInternalCorespeechtechnologyteamatAlibabacall1Ak+phonecalls,1AAk+hoursofspeechperdayAlibabasuper-apps&Taobao,Alipay,Dingding,AliOSphones/TmallTV/TmallYoukuvideoinspection(viaOurWorks–ServeExternalSpeechInteractiveVideoCourtCustomerSmartVideoCourtCustomerSmartEntryCellEntryCellInternetSpeechtech.ismanydomainsHuman-Computer TranscriptSpeechSpeechRecognition(ASR)SpeechSynthesis(TTS)VoiceRecognition(identity,gender,age,emotion)MachineLearningCoreechnologiesSpeechBigdatawithgoodcoverageofDeepSpeechadaptionSpeechMaturetechnologies,butdifficulttomakeitCoreechnologiesNaturallanguageunderstanding Data

ImageechnologiesImageBasedSearchBasedondeeplearningIndex3billionimages,latency=1AmilliondailyactiveMatchingCoreMatchingDetection+DLImageSearchDataBetterindexingEdgeEdgeModelCrossMediaSearchimagesbyRedVcoIIarshort

RedVcoIIarshortdressImagecontentFilteringout.....

......

FaceRec.IdentityVideometadata3D LowCost HighResolutionVideoechnologiesCoreVideoechnologies

…

Event/actiondetection&VideoObjectdetection,tracking,&speech,OCR,motionsVenus:VideoAnalysisIntegratedetection,tracking,recognitionvideocontentliveHighPerformance

Cloud

SoC+Edgecomputing: NaturalLanguageAliNLP:AnNLPKnowledgeSynonyms,hypernyms,Entities,Structuralizenon-structuralWordEntityDependencySentimentQuestionanswering,Machine

NLP

Content

Review

RequirementIntentionIntention

SocialMedia

Dialogue

Sentiment

Document

Text&Image

UserAddress

Language

Product

Lexical

Syntactic

Semantic

Document

Founda

C-C-

WordSeg,,POS

NER

AnalysisCorpus

News

InAliNLPKeycomponent(wordOurapproach1)state-of-the-artsystemsby1A%inE-Commerce2)industrialcompetitors(i.e.,Baidu,Tencent)by1-2%onnewsPlan:1)Outperformthebestsystemonacademiabenchmarkdataset(currently1%lower);2)Domainadaptationmodelondifferenttextgenres.AliNLPKeycomponent(sentimentWeoutperformstate-of-the-artsystems5%+inE-Commerce1-2%onbenchmarkdataset(e.g.,IMDBmoviePlan:1)Improveaspect-basedsentimentanalysisandoutperformthestate-of-the-art;2)Studytransferlearningtoadapttomultipledomains.Keycomponent

AliNLPState-of-the-artQAreliesheavilyonhandcraftedknowledgebase.Weautomaticallyextractanswersfromplaintextbydeeplearningapproaches(i.e.,machinereadingcomprehension).Plan:1)Integratelinguistic(i.e.,lexicalandsyntactic)knowledgetolearningmodels;2)CombineNLUanduserbehaviormodels;3)BuildpersonalizedQAQ:^??ùˉ|$?A:?1??AliNLPKeycomponent(structuredCrawler+scoring+humancheck,buildfive-leveltaxonomy.Useactivelearningtoiterativelylabelcontentbytag,andassociatetagtotaxonomyPlan:1)Buildfine-grainedrelationsbetweencontent,tag,andtaxonomy,andassociatethemtousers;2)Relyonunifiedvectorrepresentationofuserandcontenttobuilduser-tagIntentionGraph,anddetectuse′sreal-timeintention.axonomy: ag:DeepDeepRepresentationNeedtotrainadeeplearningmodelforeveryscenarioTwosetsofdataareusedinDataforcurrentbusinessUserbehaviordataacrossallHighcomputationalDifficulttobuildappropriate

TrainingDeepDataforcurrent

UserbehaviordataacrossallDeepRepresentationRep.

TrainingDeep

DeepFeaturesDeep

(CTR,DataDataforcurrent

Userbehaviordataacrossall

Userbehaviordataacrossall

ModelDividethelearningtaskintotwoModelDramaticallyreducethecomputationalcostwhilemaintainingsimilarModelReducemodelsize,improvecomputationalReducetherequirementofhardware,particularlyimportantforReducemodelsizeto1A%withoutaffectingpredictionImprovepredictionefficiencyby25timeson

WCC={A,1,-1,2,-

BayesianNNtocapturepredictionSuccessfullyappliedtothesingle′sday(11/11)Optimization&DecisionOptimizationis

Machine

OptimizationisfoundationofML Shipping Smart MultimediaOptimizationAssignment RoutePredictiveInventoryclick-and-mortar,drivenbyAI,improvedDecisionPrecisionAutomaticDynamicpricing,promotionDeliveryOrderDecisionDeliveryworkforceOrderRoutingManufacturingNewShiftingfrommassproductiontowardspersonalization(high-varietylow-volume,HVLV)Smallergranularity,shorterresponsetime,increasingcomplexitiesandscalesSchedulinginbridgesgapsbetweenthehigh-levelplanningandphysicalproductionDecisions:toassigntasksontoproperresourcesatpropertimeObjective:toachieveproductiongoals,suchasthroughput,lead-time,servicelevel,etc.Constraint:costpolicies,duedates,humanSchedulingoptimization:akeytotheperformanceimprovementinHVLVmanufacturing

Economyof

Predictive

Action

Predictivemaintenanceistheintelligenthealthmonitoringofequipmenttoavoidfutureequipmentfailures–Customized–Reducedmaintenance–Highersystem阿里云：阿里云資深專 蔡A1ibaba

????$[q?f三Source:??[?y±?[??——?YL????M????

Part??[?????ù?????ǘ?? ?? ?YL????M????

f[éfiv

ErasureCodingò???Part????????[??[??O???M??????êWWe(?????——4ü7üCCń???——CCń?????ňx?nPart

×?á?fl?????——

Part?????L?ū???[

ǒ—??????yGatewRoGatewRo??ǖ????

y—d??g?WebRTCIETFInternetVideoCodec商業(yè)機器智能部資深算 永???ó???apek?????àW???ó??{?p?m??j?ó?{???é??D??u?????????g??¨§b????S??????I4T??_?ìì??{U?fim??í????L??óv???D?ù??~1AK??ńD^??ó??M?}??@|??LKě? ?AI&??í?V+?fi?????fi?ê?í????b?￡§bXW?Y???w???????í?ˉ???Z??óó?H?]?l?XWY?????Ir?fi????-??L?ō?ā?tō?Teí?r?aù¤?y?????ˉ?。??ěK?????D?_ù??????????|?u?????S??o?????L??í?????Zò2AA9??īD??μμê??Q??h?J?U?????????????áT??q???????h???è?K??h???á?¨?x??í?à??h????\Vision&Beautyeaê±Q? sincez?AI?ó??ê±J-

?h????\Vision&Beautyea?h??I?x sincefi5AG?-

@¨?xsince

?h????\Vision&Beautyea?_??:?????′X?n??|ICPRMTWI2A18

Z′í????—天貓技術(shù)：營銷平臺資深算法專 志?é

??¤?,F~11fl¨

????%???&??~11L?

??1? ?é

??ē?_??3n??c?)?rfi????)???ù)lō????ōU??????d

?o)?o)

?ìIX+òg?B????-

??B???

?—

???d?????d-

holt-?é

?????~11????

?é

?ê?[?????ō(????????(+????÷??-???}?-%?C???&???¤???í??搜索事業(yè)部：AI@資深算法專 三[д,NW;q=??u8[h-

1ur\[0Y

zN*3品MIT'-,-.%201610(&'/LL4),0/LL#*)$11m???)aFS/(reinforcementMulti–agentreinforcementlearning:3品A)3品B7-;agent??c4action?—agentS??c?AbM,?—

Actor

Actor

:actor,+critic,actorQ)LSTM,Q)'2uUo?1c)Ec?h-2uUo?1c)Ec?h-~]?he???–DeepUserPerception10B2uUo?1c)Ec?h-~]?he???–DeepUserPerception3φ @ ?3!+"A%*4uY?N>?品 ?? ? ?????

???????4uY?N>?品?–DeepGraph;AQ;(&?Q4uY?N>?品?–DeepGraph55 =

(

--------…-StandardMatrixVectorProduct-…-…-)*

---------------

Multi-bitBinary ModifiedMulti-bitBinaryf αibi) {αi,

=arg w

α′ 2.

{α',b'--1--1…-1…-1…---1…-1…---1

??a￡?￠(NLP),?0品2A:品資深技術(shù)專 粵e?

ü?:rerrerSupplier

RetailerRetailerRetailerRetailerRetailerSupplierRetailerSupplierRetailerrer

FCAFCBSupplier

SCSCSC rerSCrer

LMDepotALMDepotBLMDepot

ConsumerAConsumerBConsumerConsumerRetailerConsumerRetailerRetailerRawMaterialSupplyRetailer

B2C

SupplyChainSupplierSupplySupplierSupplyX§??_??μμn:

Return&

Localcourier)

B2C(LMDelivery/Localcourier)BrickandMortar

Self-

Key Demand Key Demand

X§??X??&X§???:

????-?y?d:

c(i):argminIIx(i)

1{c(i)j}x(iuj

1{c(i)

????-p※?Oúí:

??(?O?1ú) ?í?O KVI,KVCKVI,KVC

??—

O??—

?à???a:

X???-:gēě?a:

X???-?o&|?¤?: X§?P?:X§?ā??—v@X§?Y????,?R-??@X§??<y?`LY?,D??R-?<āX§?Y???TX§??U-WorkflowinSales

Inv. Order SupplyChainAlg.Demand FC Safty Inv.

Inv.

Fulfillment

OrderWorkflowinSales Inv. Order

????-?K?ìbú:

1B

2?&

ò:@J??VS.螞蟻金服資深技術(shù)專 南Part?????2A14?1A???????_ü?|?ES??ü??N????G}????E???üEo?Dd?ü??E?§ā\??Eoǚù????ǜ?I?K?????Ví?o????????????@???·ü?hOES???????Eè·±??ˉfi???l??????w????????????§T￡?

Techfin2AA4- 2AA7- 2A1A- 2A13- 2A16-Part?d???é???:??????

inPart?d?|???

Theimpossibilitiestodayleadtothebreakthroughs???oǒ$,?ǒ$?ó?dě?·???ī3?d?|????????r????99.99%

6K

I?K???n??d?|??????????d?|????????-fl??ǚ??zgǖ???d?|???????

AI

AIPart?d????????????

?d???????????|t??|èūAA?J?éTOP-?d????????aó??(?fi)-

?d???????

?d???y?ê?

?d???y?ê?–ò?ó?d???Antone ???—?d???Antone??￡?

?d?????????……

?d?????????i???ō?????Lōá???)???A/BTesting?y ???????′k? ??????? ?????? ??ì?Part???????′ǐ???·~uy+

?v¤§??′

???ń?: ???ń?:??????

Part????@???? ?·t¤??:smalldata,searchspace ???,Decision????ú¤÷?

Offlinegraphgraphupdatin,onlinedecisionmaking?ō?

ABWhat-ifscenario????????????

DecisionProcess?????ō?

BigData&ForecastingProblem

Multi-agent

NLP& ??t¤??éX???n??b???_ Part????????

?????

?ī???ī??

?p???? ??fi??[??h?:?GQHQFQID5r5/481tn?~fi????Q??aμ??MLEx???-￡?ǒ?????–?1—?$????ü`??,????????￡?z$,ò?￡?ǒ???(ETA,EstimateTimeof??:???g??+ Pre-Span,KNN,LR,RNN,

RNN,LSTM,???āt¤???āt¤–Outpu:àg?¨·??—–?—–GeneralizedVehicleRouting–VRP–Clustering+localMulti-agentPathFinding(MAPF):??|?ā?????—fi?[??i??—p:Makespan?CumulativeTimeCost()

Multi-agentPathFinding(MAPF):??|?ā?????—fi?[??i??—p:Makespan?CumulativeTimeCost()

Makespan:8????:????三???????:—

?ǘù??(maxRectangle,extremePoint,?gˉ??三?????Reward

Node?y???:state,parent,child,visit_num?????3DBinPackingDerivativeFreeDynamicProgramming+convexhull???ā???g?ā??

VehicleRouting ???? ?g?