PLE — 一種新的分層萃取多任務學習網絡結構.pdf

1、Progressive Layered Extraction (PLE）Output AOutput B一種新的分層突取多任務不個Tower ATower B學習（MTL）模型和由此關于AdI的一點思著劉軍寧InputTencent騰訊#page#Short Bio廣秀淘Tencent騰訊E螞蟻金服新華智云一OnBDADVERPLEXTencent騰訊#page#Outline1.Background: Recommender System GRS）MultitaskLearning（MTL）多任務學習為什么MTL成為RS廣泛采用范式2.PLE分層萃取多任務學習模型3.強人工智能-AGI的一點

2、思考Tencent騰訊#page#BackgroundRecommender System: Most widely used AI technique2勺視頻信息流：YouTube微視19BIGOLIVE花椒直播淘寶直播抖音快手下一代信息流TikTokQO視頻號meitu美圖社交：學看一看淘蘑菇街美團電商：woouozeue京東頭條傳統信息流：騰訊新聞知乎新浪微博今日頭條騰訊看點第一代信息流Tencent騰訊#page#BackgroundIndustry RS Mainstream Approach: Multitask learning(MTL) 排序層9.工業界案例分享多目標學習在推薦

3、系統中的應用（1）美團本文櫥路美團消你喜歡”深度學習排序模型實踐，地址R（2）知乎進擊的下一代推薦系統：多目標學習如何讓知學用戶互動率提升100%？，地址：保2020 Google開發者大會33s/GUMz官方最全技術HjQvzdGVOkKz4zA口干貨集錦?。?）美面0馬上直收當推薦退到社交：美面的推薦算法設計優化實踐，地址：nES多任務學習在美圍維薦排序的近期實踐，地址QOSW33Q8QTfATTencent騰訊#page#Common MTL StructuresHard Parameter Sharing Caruana， 1997Task ATask Bsame sharing pa

4、rameters for different tasksTask-specificlayers=negative transferTower ATower BSharedlayersInputTencent騰訊#page#Common MTL StructuresHard Parameter Sharing ICaruana,1997same sharing parameters for different tasksnegative transferASKs Soft Parameter Sharing between single task and hard sharingConstrai

5、nedlayersTencent騰訊#page#Common MTL StructuresCross-stitch Network Misra et al，2016Hard Parameter Sharing ICaruana,1997 same sharing parameters for different tasksnegative transfer Soft Parameter Sharing between single task and hard sharingCross-stitchNetwork（十字繡網絡）IMisraetal，2016&SluiceNetwork（水閘網絡）

6、Ruderetal，Sluice Network Ruder et al，20172017static fusing weights售Tencent騰訊#page#MTL Benefit增強遷移/泛化：InductiveBias歸納偏置Knowledgesharing：InductiveTransfer歸納遷移，減少訓練數據量需求BetterRegularization鼓勵模型學通用的共享的東西利于學到適用新任務上的泛化表征Coordinated representation learning互助特征表征學習- More efficient representation learning co

7、mpared to single taskGeneral Learning FrameworkNLPCV Recommendation SystemTencent騰訊#page#Why widely adopted in RS？用戶滿意度難以直接學習：非常復雜的深度語義變量無直接pointwise量化監督S月14日度快辦證哈利波特整體對應人均時長+留存率豐富具體隱式反饋：目標偏差|物品偏差|用戶偏差subscribe充分學習，用上所有掌握的信息helpsrep.learningMTL高效協同學習具體任務play/pauseMTL工程便利-like為什么是最近2年：傳統feeds流產品形態全屏沉

8、浸式詳情頁comnment傳統相對串行：Ctr/Cwr微視額百萬紅傳統模型結構的不斷換代shareLR/GBDT/deep&wide/FM/DNN/DeepFM哈利波特與發布超清修復版favor.etc照，時隔188月14號重映強化學習RL進展navigateTencent騰訊#page#Multi-goal Fusion Ranking in RS通過樣本權重進行多目標優化簡單粗糙多個單模型分數融合：多目標人工上帝拍定融合公式：聘子摸象長期目標監督融合公式學習還是Learning To Rank (LTR）：單目標？人工heuristic定義多種規則偏序關系偏序關系難以完整覆蓋和精確定義，監

9、督不充分MTL +Rerank:MTL in RS: hard sharing ESMMMMOEPLETISH sokea uorees pyx6:reIogTencent騰訊#page#ESIMM（Entire Space Multitask Model）PCTCVR阿里18年提出2CTF全在中主要解決CTR/CVR類型不對稱串行任務緩解SSB（SampleSelectionBias）Concatenate緩解DS（DataSparsity）Field-wisePoolng Layer局限EmbeddingCVR仍可更充分監督Laye核心在于保證Conversion/Impressionus

10、er tielcAssumption:CVR-tCTR-tasl最大化Pointvvise轉化=最大化用戶體驗L）（y，f（x0）+（y&zf（x0）xf（x0）Tencent騰訊#page#MMoE（Multi-gate Mixture-of-Experts）OutrCuiOutput AOutput BglamolTowerATowerATowerBGaleAGalaBExpertoExpertExpenGateExpert0Expent1Expent2MoEMMOE Ma et al，2018 Google提出Bagging式的多專家集成no task-specific expertCo

11、nditional Computation:single-level experts and gatessingle gate for alf tasksTencent騰訊#page#Challenge - seesaw phenomenonComplex task correlations Negative Transfer： Seesaw phenomenon：improve performance of one task often at the price of others0.130Complicatedly Correlated Task Group1 watchtime 2 th

12、res0.131VTR（View Through Rate）10watch time=50K）Task2（marital status =never married）MMOEPLEModelsAUCMTLGainonv.MTLGainCorrelation 0.2Correlation 0.4Correlation 0.510.010-0.9923Single-Task0.94450.0050.005SoooMMOE0.9393+0.00480.9928+0.0005PLE0.9522+0.00780.9945+0.00220.000oo0o0.000Task1Task2Task1Task2T

13、ask2Task1Ali-CCP DatasetCorrelationAverage0.005（84M samples fromn Taobaos recommender system）0.000.005CTRClick-Through Rate）CVRConversion Rate）Models-0.01oooo0.000AUCMTLGainAUCMTLGainTask1Task1Task2Task2Task1 Task2 overallSingle-Task0.60880.6040MMOE0.6094+0.00060.5738-0.0302Vconsistently outperforms

14、 across differentPLE0.61120.6097+0.0024+0.0057correlation patterns and different applicationsTencent騰訊#page#Contribution of PLEseparates shared experts and task-specific experts explicitly= introduces multi-level customized gate control with an novel progressive separationroutingsCoordinated represe

15、ntation learning共享學習網絡共享學習網絡共享學習網絡（結構分化，浙進分離）（分解但無分化）協同表征共享學習簡單輸出共享學習outperforms state-of-the-art MTL models， eliminates the negative transfer.and seesawphenomenon across different task correlations，task-group sizes and applicationsProgressive Layered Extraction CPLE):A Novel Multi-Task Learning (MT

16、L) Model forPersonalized Recommendations hps:ildl.acm.org/doilaba/l0.1145/3383313.3412236ACMRecSys2020最佳長論文（Best LongPaperAward）Tencent騰訊#page#NAS in MTL Still Early StageStrong assumptionnarrow framework= Network structures are not continuous: both structure and utilityHard to find gradient pathTen

17、cent騰訊#page#AGI的一種設想Task 3Task 1New TaskTask 2表達層Meta learning推演層MTLChoose modulesfor each layer表示層MTL輸入層Tencent騰訊#page#AGI的一種設想Learning Layers輸人層：Quantization&Normalization表示層（MTL）推演歸納層（MTL）表達層C Standardized module libraries for each layer開放工具箱，可進化MetaLearning 模塊 choose modules for each layer000判斷設計新網絡的必要性工具擴展進化：GAN，ES，RLTencent騰訊#page#謝謝哈聽！Tencent騰訊