微博机器学习与深度学习在推荐系统的实践

1 .u u 0 IE2 m & %- ) 

2 .• • • u • u f • • • • • o 

3 .1 931B9> u u 

4 .% • 

5 .% • k 

6 .& • • kf q f • h • • h u • h u • h u 

7 . u • w • h qf • h f 

8 . u WMN WMN B S input WMN WMN : OS B S WeiFlow process WMN B S B S process WMN B S output WMN N ZRL 

9 .u z q : OS q q 8MO q 

10 . u • p • fz f f • rT • f • f • : OS f • rT t • • • rT o • l • 3 NLSXWRV 

11 . u • l iu m I 1 -/ % % % % & @ 6 & C46 & & 4NNX6 C9 81A8 A74 6B@ ( 1M PZ M 

12 . u • m • @. f • 6 h k • 4NNX6 • • 6B@ h u f f K L • 1M PZ M . u f f • A74h f f u • o • 8 h2:4@ 3R 8 f94 

13 . u Predict Score A D A D A D A D NZ RVPModel Serving System server serverserver serverserver Serving PS server server server Traing PS Sample Data worker worker worker workerworkerworker worker Traing Model System 

14 . u Model Serving System PSproxy PSproxy HA Fault tolerance checkpoint PSscheduler zookeeper checkpoint PSserver PSserver PSserver File System Param Server System Local PSagent PSagent HDFS PSsubmit Model Training System Model register Status set/get Model delete Model Save Model Load 

15 . u • c • f • m • h f • h H: f f 21A5 f l • l h Al A f • v • h X L NM TNZ l f • h L NLSXWRV WL T @NUW N f f f • • h C CA8 f f f ) % • h l @C f 73 f ) • h d@1 75 81A8 >4ef f & ) • h s 2 @W_ 2 :N f ZW_ f - 

16 .u m m m l NZ WV% 1C3 b m NZ WV& b b b C3 B A AC3 B 

17 . u • m R • • • l f • • : A MNXTW UNV f l • H: f l • • h l f • h AA5g1DF 

18 . u • m • 6 h f • f f fn f • f f d f e • f • f ) •l h m • 6 6 f ) • • 6 6 f • 6 h 6 

19 .( u • • h NUKNMMRVP • h u f ERMN 4NNXi4NNX6 h9 NU& NL h @ /E 4 /6 /4NNX6 h6 66 4AA h 6 z 6 h49 495 B4 h4NNX6 RZ ER N @ VS W g g me a p 9 NU& NL 4AA 6 66 z @ E 4 o f r rl q 6 4NNX6 z f f q 4NNX6 f q 2 @ 

20 .( u u • u m I1 2 1 2 u m o . AX ZS ERMN 4NNX 4NNX6 3B@ BNV WZOTW_ 4NNX 3ZW N _WZS + - k l AB 7@C RVLNX RWV ZN VN 25@B D77 GWTW 

21 .( u u • • Ah2A AA 1A • 9 @RVP1TTZNM LNh8WZW WMf 9 PZXLf i f i • A 9h4R ZRK RWVA Z NP 1 9f sf l • 6 % f 6 & f • • f l 9 5 9 5f l9> W NZT X 

22 .( u • m I1 / D m psS DNN/DeepFM/W&D LR/GBDT CNN m Embedding m / / / RPC kubernetes/ol-submit CPU GPU 

23 .( u • • h WXNZ WZ O RWV F 1 BD XZ VN OTW % Y V Ra RWV • h K L RVP BNV WZ@B A AA5 1DF NWV • 2B CA • • NV WZOTW_ • 32B • 7 Cf 

24 .( u • • 5UKNMMRVP f f • _WZM& NL KNZ NUKNMMRVP f v s f v • r 5UKNMMRVP v f v • C NZ 9 NU 5UKNMMRVP • 9 NU& NL 6 f k • • m • 4NNX6 6 f ( • ERMN 4NNX @ f ) • r 4NNX • 

25 . 1B6>@ f 

26 .% • f f w i N O 

27 .& / / m n /IDE tW 

28 .& 6NNM bb iu f F1 2B2 mF1 E n F1 p tW 8 MWWX AX ZS A WZU 6TRVS BNV WZ6TW_ 3 OON bb 8MO >MX A ENR A G ZV : 4WLSNZ y 7 C bb 

29 . Feed / /z b iu f o f19-22 f h z tf f h z …… 9 f f f f w f …… WeiFlow p f…… h Y=f(x1,x2……,xn) f h WeiIDE f…… k