AiCon 全球人工智能与机器学习技术大会

深度树匹配 下一代推荐技术的探索和实践

1. ——
2. 2013
3. • • —— • • ——
5. 推荐召回问题的计算限制 点计算消耗 所需计算次数 系统性能边界 TopK
6. youtube 1 1 Item Embedding(IE) cm! Imk! 2 2 IE ux! 3 item 3 User Embedding UE UE K IE
7. —— ==> +
8. • • —— • • ——
9. + 单点计算消耗 =>T 所需计算次数 + =>N != => 系统 能边界
10. => Tree-Based Deep Match TDM 10 10 Top1 ->30 1 2 3
11. RN LN1 topK BeamSearch —— topK —— topK —— SN1 LN2 K=2 SN2 SN3 SN4 ITEM1 ITEM2 ITEM3 ITEM4 ITEM5 ITEM6 ITEM7 ITEM8 O(2*logN*K) N K
12. —— ρln2 = MAX(ρsn3 , ρsn4 )/α1 α1= Σ(ρln1 , ρln2) RN LN1 LN2 ρsn4 = MAX(ρi7 , ρi8 )/α2 α2= Σ(ρsn1 , ρsn2, ρsn3, ρsn4) SN1 TopK SN2 SN3 SN4 ρit8 = P(ITEM8 user) TopK ITEM1 ITEM2 ITEM3 ITEM4 ITEM5 ITEM6 ITEM7 ITEM8 user BeamSearch TopK TopK
13. RN! LN1! SN1! ITEM1! 1 2 Topk LN2! ! SN2! ITEM2 ITEM3! ITEM4! SN3! ITEM5! ITEM6! SN4! ITEM7! BehavedLeaf ITEM8!
14. 1 Interest Network 2 Cross info Deep
15. RN LN1 LN2 Item SN1 SN2 SN3 SN4 … ITEM1 ITEM2 ITEM3 ITEM4 ITEM5 ITEM6 ITEM7 ITEM8 == == X min L(X,Y )⇠T (X, Y, ✓) item pair Recall + 53% X min L(T ) = i,j2S distT (i, j) Y ( )
16. 1 ——BeamSearch 2 —— 3 —— 1 2
17. TDM TDM TDM SIGKDD2018 Method1 Item-CF2 Youtube VecSearch3 TDM RPM Shop/ 1 2 3 TDM ETREC Youtube Recall 6.95% 7.58% 12.37% 20%+ Ad 80%+ UserBehavior Lift 9.06% 77.98% New Category 1.06% 3.09% 4.82% Lift 191.51% 354.72% !
18. TDM + UTS+DS+MS & KW —— —— TOP2K pv gpuFLOP 50 RT 5%
19. TDM TopK RN! LN1! —— + SN1! ITEM1! + LN2! ! SN2! ITEM2 ITEM3! ITEM4! SN3! ITEM5! ITEM6! SN4! ITEM7! ITEM8!
20. • • —— • • ——
21. + + end2end
22. Recall min L(X,Y )⇠T (X, Y, ✓)
23. ecpm pctr*bid ecpm RECALL ecpm ecpm RN LN1 = ecpm ecpm = SN1 LN2 SN2 SN3 SN4 ITEM1 ITEM2 ITEM3 ITEM4 ITEM5 ITEM6 ITEM7 ITEM8
24. User RN SEARCH-POLICY LN1 SN1 Model Train LN2 SN2 SN3 SN4 Match Serving ITEM1 ITEM2 ITEM3 ITEM4 ITEM5 ITEM6 ITEM7 ITEM8 LAYER-RANKING-MODEL Probability Softmax (2) PReLU & BN (24) PReLU & BN (64) User Co-train PReLU & BN (128) Retrieval Concat Time Window N Scores Time Window 2 Time Window 1 Weighted Average X Nodes X Item 1 Weight X Activation Unit Activation Unit Embedding Layer Item 1 Item N Weight Item 2 Weight … … … … Item 2 Activation Unit Node Item n User Behaviors RN LEARNED-INDEX LN1 SN1 Behaviors LN2 SN2 SN3 Items SN4 ITEM1 ITEM2 ITEM3 ITEM4 ITEM5 ITEM6 ITEM7 ITEM8 Offline Online Items item item RN ( sideinfo! ( ( ( RN RN ( () ,( RN RN itemb itema
25. • • —— • • ——
26. TDM TopK
27. TDM User Page View! Request with ! User Context! TDM Ad Candidates! (Tens of Thousands)! ! Page View! Response of Ads! (Tens)! TDM! Rank (CTR Predict) Ad Candidates! (Hundreds)! Strategy ! !
28. “ ” 2018.12( ) 2019.03( TDM二期 ) 完成简单在线Serving系LD开 源,支持基础树检索 TDM一期 TDM三期 SXD 一起开源,支持离线训 练和T测 丰富和完善在线Serving系LD开源, 具备支持十亿规模超大候选D检索M力 2019.01( ) XDL https://github.com/alibaba/x-deeplearning
30. Q&A