Diese Präsentation wurde erfolgreich gemeldet.
Wir verwenden Ihre LinkedIn Profilangaben und Informationen zu Ihren Aktivitäten, um Anzeigen zu personalisieren und Ihnen relevantere Inhalte anzuzeigen. Sie können Ihre Anzeigeneinstellungen jederzeit ändern.
PyCon Korea 2019
, .
|
1
2
/
+
?
?
5
?
6
?
7
?
8
?
?
.
14
( )
?
?
15
10
20
20
16
.
17
?
.
- -
.
- -
?
22
( )
?
?
:
23
( )
?
?
( ) .
24
10
20
20
: .
25
10
20
20
20
26
: .
27
.
?
.
30
/
TP.
.
.
!32
33
…
34
- - ViewerEnd
• : 

CTR(%) 

• 

• MAB(Multi Armed Bandit) 

• User Clustering
-
!35
MAB(Multi Armed Bandit)
• MAB = Exploration( ) and Exploitation( ) Trade-off

• 10%( ) Feedback (impression,
click)
* ε-gre...
• Feedback CTR(%) . 

• CTR(%) = # of clicks / # of impressions
Exploration
0.4% 4.0% 2.9% 7.3% 2.7% 8.7% 6.7% 1.0% 1.9% 8...
• CTR 90% ( ) 

• CTR
0.4% 4.0% 2.9% 7.3% 2.7% 8.7% 6.7% 1.0% 1.9% 8.1% 6.3%
3.6% 6.7% 8.0% 3.1% 3.6% 2.0% 4.4% 3.1% 7.3% ...
0.4% 4.0% 2.9% 7.3% 2.7% 8.7% 6.7% 1.0% 1.9% 8.1% 6.3%
3.6% 6.7% 8.0% 3.1% 3.6% 2.0% 4.4% 3.1% 7.3% 8.2% 2.7%
4.4% 8.1% 0....
0.4% 4.0% 2.9% 7.3% 2.7% 8.7% 6.7% 1.0% 1.9% 8.1% 6.3%
3.6% 6.7% 8.0% 3.1% 3.6% 2.0% 4.4% 3.1% 7.3% 8.2% 2.7%
4.4% 8.1% 0....
ε-Greedy MAB ε=0.10
41
10M Impressions
10%(ε)
1M Impressions( )
1 2 3 4 4 5
6 7 8 9 … 100
ε-Greedy MAB ε=0.10
42
10M Impressions
10%(ε)
1M Impressions( )
1.1% 2.0% 8.2% 0.01% 4.6% 1.2%
5.2% 0.1% 0.2% 1.0% … 2.2%
...
ε-Greedy MAB ε=0.10
43
10M Impressions
10%(ε)
1M Impressions( )
1.1% 2.0% 8.2% 0.01% 4.6% 1.2%
5.2% 0.1% 0.2% 1.0% … 2.2%
...
ε-Greedy MAB ε=0.10
44
10M Impressions
10%(ε)
1M Impressions( )
1.1% 2.0% 8.2% 0.01% 4.6% 1.2%
5.2% 0.1% 0.2% 1.0% … 2.2%
...
ε-Greedy MAB ε=0.10
45
10M Impressions
10%(ε)
1M Impressions( )
1.1% 2.0% 8.2% 0.01% 4.6% 1.2%
5.2% 1.5% 0.2% 1.0% … 2.2%
...
ε-Greedy MAB ε=0.10
46
10M Impressions
10%(ε)
1M Impressions( )
1.1% 2.0% 8.2% 0.01% 4.6% 1.2%
5.2% 1.5% 0.2%
10 Impressio...
ε-Greedy MAB ε=0.10
47
10M Impressions
10%(ε)
1M Impressions( )
1.1% 2.0% 8.2% 0.01% 4.6% 1.2%
5.2% 1.5% 0.2% 1.0% … 2.2%
...
ε-Greedy MAB ε=0.10
48
10M Impressions
10%(ε)
1M Impressions( )
1.1% 2.0% 8.2% 0.01% 4.6% 1.2%
5.2% 1.5% 0.2% 1.0% … 2.2%
...
ε-Greedy MAB ε=0.10
49
10M Impressions
10%(ε)
1.1% 2.0%
5.2% 1.5% 0.2% 1.0% … 2.2%
Best
arm
( )
90%(1-ε)
(100 ) 10k Impres...
Thompson Sampling MAB ?
Thompson Sampling MAB
• (arm) CTR Beta(a,b) . ( a=click, b=unclick )
51
1
10%
Impressions : 10 50 100 200 1k 10k
2
25%
Imp...
Thompson Sampling MAB
• (arm) CTR Beta(a,b) ( a=click, b=unclick )
52
1
10%
( ) CTR 15%
1 (10%<15%) 100 Impressions trial
...
Thompson Sampling MAB
• (arm) CTR Beta(a,b) ( a=click, b=unclick )
53
2
25%
2 CTR 25%>15%
( )
.
Impressions : 10 50 100 20...
• 1K Impressions
54


CTR
• 10K Impressions
55


CTR
• 50K Impressions
56


CTR
• 100K Impressions
57


CTR
• 500K Impressions
58


CTR
• 1M Impressions
59


CTR
• 1M Impressions
60


CTR
CTR (Arm)
• 1M Impressions
61


CTR
CTR
( )
.
.
• 1M Impressions
62


CTR
CTR
( )
.
.
TS-MAB
& Trade-off
Regret( ) .
User Clustering
• CTR .
CTR : 7.6%
A : 25% (30 )
B : 2.1% ( )
C : 7.1% ( )
CTR
!63
User Clustering
• X CTR 

• CTR
200 8,000
User Clustering
• 

• 8
8
8,000 64,000
CTR
Clustering ?
CB(image,Text)
Feature User Feature
[0.628, 0.88, 0.376, 0.065, 0.849]
[0.508, 0.268, 0.193, 0.125, 0.425]
[0...
Clustering ?
14 CB(image,Text)
Feature User Feature
[0.628, 0.88, 0.376, 0.065, 0.849]
[0.508, 0.268, 0.193, 0.125, 0.425]...
-
• 

• #1, #5, #6, #7 

• #0, #2 #3
-
• 

• #1, #5, #6, #7 

• #0, #2 #3
/
#1
, ,
#3
, ,
/
#3
, ,
Tag
Tag
/
.
?
77
#2
#1
#3
User Clustering
Targeting
CTR
1 : CTR 9.1%
2 : CTR 8.8%
3 : CTR 8.0%
4 : CTR 7.8%
5 : CTR 7.1%
6 : CTR 6.8%
...
?
78
#2
#3
User Clustering
Targeting
Ranker
MAB
Ranking
Targeter
+ MAB
-ViewerEnd
• : 

CTR(%) 

• 

• (Item Feature)

• MAB(Multi Armed Bandit)
!80
?
81
#2
#3
User Clustering
Targeting
Ranker
MAB
Ranking
Targeter
?
82
(Item)
Targeter
Feature
Targeting
Ranker
MAB
Ranking
-ViewerEnd
(CF)
(Text)
(Image)
!83
1 2 3
- Item Features
/ Image 

(1) Image Feature
, Text 

(2) Text Feature
Feedback


(3) CF-Feature
!84
?
!85
?
!86
?
!87
1 3 41 1 92 1
1 3 41 1 92 1
1 3 41 1 92 1
Image
Text( )
CF( )
!88
1 3 41 1 92 1Image Style
Image
Text( )
CF( )
Image Style
Style transfer network
!89
1 3 41 1 92 1
Image Style
Text( )
CF( )
Image Object Detection Task
Image Object
Pre-trained VGG19 Model
!90
1 3 41 1 92 1
Image Style
Image
CF( )
,
Keyword
Word2Vec
!91
1 3 41 1 92 1
Image Style
Image
Text( )
CF( )
Matrix Factorization(ALS)
with implicit feedback
(Feedback) Item-User
!92
1 3 41 1 92 1
!93
1 3 41 1 92 1
1 3 41 1 92 1
1 3 41 1 92 1
1 3 41 1 92 1
“ ”
(%)
!94
0.4% 4.0% 2.9% 7.3% 2.7% 8.7% 6.7% 1.0% 1.9% 8.1%
0.4% 4.0% 2.9% 7.3% 2.7% 8.7% 6.7% 1.0% 1.9% 8.1%
0.4% 4.4% 2.9% 7.3% 2....
0.4% 4.0% 2.9% 7.3% 2.7% 8.7% 6.7% 1.0% 1.9% 8.1%
0.4% 4.0% 2.9% 7.3% 2.7% 8.7% 6.7% 1.0% 1.9% 8.1%
0.4% 4.4% 2.9% 7.3% 2....
?
97
(Item)
Targeting
CTR
1 : CTR 9.1%
2 : CTR 8.8%
3 : CTR 8.0%
4 : CTR 7.8%
5 : CTR 7.1%
6 : CTR 6.8%
7 : CTR 6.7%
…
MAB...
+ MAB
!98
!99
100
1 2 3 4 5 6
…
89 90
(Clicks)
Impression
101
1 2 3 4 5 6
…
89 90
(Use Coin)
Impression
102
User Cluster +
+
MAB
MAB
RankerTargeter
=
=
103
+
MAB
Conditional Bandit
Exponential Smoothing
Seen Decay
Soft User Clustering
Retention Model
Unbiased Most Popular
F...
104
• MAB

- Bandit Algorithm = Thompson Sampling(

- Reward = Click (with Unclick )

- Play Arms = Cluster Most Popular 
...
105
• MAB

- Bandit Algorithm = Thompson Sampling

- Reward = Click (with Unclick )

- Play Arms = Cluster Most Popular 

...
106
1. MAB ?
(Beta) (Alpha) -20% —>
? ?
MAB ?
-20%
2. Conditional Bandit
107
1 2 3 4 5 6
…
89 90
by @troye.kwon
2. Conditional Bandit
108
1 2 3 4 5 6
…
89 90
Impressions
Reward=Click( )
α=click, β=unclick
MAB
by @troye.kwon
2. Conditional Bandit
109
1 2 3 4 5 6
…
89
Impressions
Reward=Click( )
α=click, β=unclick
MAB
Reward=Use Coin( )
α=use-coi...
2. Conditional Bandit ?
110
(Beta) (alpha)
? ?
- MAB .
3. Retention Model
• : . 

, 

“ ” . 

• 

• MAB
111
by @jinny.k
+
MAB
Targeter Ranker
3. Retention Model ?
112
by @jinny.k
(CTR) (CVR)
? (CTR) ?
4. Seen decay
• : Negative Feedback 

• click impression Ranker 

• : (alpha) (Beta)
113
-> CTR
114
• 

• Hard Clustering(k-Means) —> Soft Clustering (pLSI)

• Feature Matching 

• Targetting Genre/Tag Matching ...
(%)
!115
Soft Clustering (pLSI) Feature Matching
Conditional Bandit Retention Model
Exponential Smoothing
MABUnbiased Most...
?
?
!116
!117
?
118
?
119
?
120
?
121
?
122
0.1%
123
= 3.96%
= 6.70%
= 2.63%
= 0.07%
-> 4.56
124
= 3.96%
= 6.70%
= 2.63%
= 0.07%
4.56
-> 4.56
125
= 3.96%
= 6.70%
= 2.63%
= 0.07%
4.56
AB
(<0.001)
Feedback (>4.56days)
/
,
?
.
127
Base : Editor’s ( X) 1.9%
Alpha : 1 4.8%
Beta : 2 5.5%
Gamma : 3 6.5%
CTR
1 2 3 4
…
.
128
Base : Editor’s ( X) 1.9%
Alpha : 1 4.8%
Beta : 2 5.5%
Gamma : 3 6.5%
CTR
+ 242% + 42%
1 2 3 4
…
CTR
.
?
/ ?
?
/ ? ->
? ->
133
YOU
?
.
!134
|
추천시스템 이제는 돈이 되어야 한다.
추천시스템 이제는 돈이 되어야 한다.
추천시스템 이제는 돈이 되어야 한다.
추천시스템 이제는 돈이 되어야 한다.
추천시스템 이제는 돈이 되어야 한다.
Nächste SlideShare
Wird geladen in …5
×

추천시스템 이제는 돈이 되어야 한다.

7.807 Aufrufe

Veröffentlicht am

Pycon2019에서 발표한 자료 입니다

Veröffentlicht in: Daten & Analysen
  • Als Erste(r) kommentieren

추천시스템 이제는 돈이 되어야 한다.

  1. 1. PyCon Korea 2019 , . | 1
  2. 2. 2 / +
  3. 3. ?
  4. 4. ? 5
  5. 5. ? 6
  6. 6. ? 7
  7. 7. ? 8
  8. 8. ?
  9. 9. ?
  10. 10. .
  11. 11. 14 ( ) ? ?
  12. 12. 15 10 20 20
  13. 13. 16
  14. 14. . 17
  15. 15. ?
  16. 16. . - -
  17. 17. . - -
  18. 18. ?
  19. 19. 22 ( ) ? ?
  20. 20. : 23 ( ) ? ? ( ) .
  21. 21. 24 10 20 20
  22. 22. : . 25 10 20 20 20
  23. 23. 26
  24. 24. : . 27
  25. 25. .
  26. 26. ?
  27. 27. . 30 / TP.
  28. 28. .
  29. 29. . !32
  30. 30. 33 …
  31. 31. 34 - - ViewerEnd
  32. 32. • : 
 CTR(%) 
 • • MAB(Multi Armed Bandit) • User Clustering - !35
  33. 33. MAB(Multi Armed Bandit) • MAB = Exploration( ) and Exploitation( ) Trade-off • 10%( ) Feedback (impression, click) * ε-greedy MAB . !36
  34. 34. • Feedback CTR(%) . • CTR(%) = # of clicks / # of impressions Exploration 0.4% 4.0% 2.9% 7.3% 2.7% 8.7% 6.7% 1.0% 1.9% 8.1% 6.3% 3.6% 6.7% 8.0% 3.1% 3.6% 2.0% 4.4% 3.1% 7.3% 8.2% 2.7% 4.4% 8.1% 0.6% 5.9% 9.2% 7.3% 8.3% 8.6% 4.2% 9.9% 6.9% * ε-greedy MAB . !37 MAB(Multi Armed Bandit)
  35. 35. • CTR 90% ( ) • CTR 0.4% 4.0% 2.9% 7.3% 2.7% 8.7% 6.7% 1.0% 1.9% 8.1% 6.3% 3.6% 6.7% 8.0% 3.1% 3.6% 2.0% 4.4% 3.1% 7.3% 8.2% 2.7% 4.4% 8.1% 0.6% 5.9% 9.2% 7.3% 8.3% 8.6% 4.2% 9.9% 6.9% Exploitation8.0% 8.2% * ε-greedy MAB . !38 MAB(Multi Armed Bandit)
  36. 36. 0.4% 4.0% 2.9% 7.3% 2.7% 8.7% 6.7% 1.0% 1.9% 8.1% 6.3% 3.6% 6.7% 8.0% 3.1% 3.6% 2.0% 4.4% 3.1% 7.3% 8.2% 2.7% 4.4% 8.1% 0.6% 5.9% 9.2% 7.3% 8.3% 8.6% 4.2% 9.9% 6.9% Exploitation (10%) (90%) & ? : ? : . !39 MAB(Multi Armed Bandit)
  37. 37. 0.4% 4.0% 2.9% 7.3% 2.7% 8.7% 6.7% 1.0% 1.9% 8.1% 6.3% 3.6% 6.7% 8.0% 3.1% 3.6% 2.0% 4.4% 3.1% 7.3% 8.2% 2.7% 4.4% 8.1% 0.6% 5.9% 9.2% 7.3% 8.3% 8.6% 4.2% 9.9% 6.9% Exploitation Exploration(10%) Exploitation(90%) • MAB ? • TS-MAB ε-greedy UCB(Upper Confidence Bound) Lin-UCB Thompson Sampling NeuralBandit LinRel (Linear Associative Reinforcement Learning)  !40 MAB(Multi Armed Bandit)
  38. 38. ε-Greedy MAB ε=0.10 41 10M Impressions 10%(ε) 1M Impressions( ) 1 2 3 4 4 5 6 7 8 9 … 100
  39. 39. ε-Greedy MAB ε=0.10 42 10M Impressions 10%(ε) 1M Impressions( ) 1.1% 2.0% 8.2% 0.01% 4.6% 1.2% 5.2% 0.1% 0.2% 1.0% … 2.2% (100 ) 10k Impression
  40. 40. ε-Greedy MAB ε=0.10 43 10M Impressions 10%(ε) 1M Impressions( ) 1.1% 2.0% 8.2% 0.01% 4.6% 1.2% 5.2% 0.1% 0.2% 1.0% … 2.2% (100 ) 10k Impression CTR = 1.5% Best arm ( ) 3 8.2% 7 5.2% 4 4.6% 50 3.0%
  41. 41. ε-Greedy MAB ε=0.10 44 10M Impressions 10%(ε) 1M Impressions( ) 1.1% 2.0% 8.2% 0.01% 4.6% 1.2% 5.2% 0.1% 0.2% 1.0% … 2.2% Best arm ( ) 3 8.2% 7 5.2% 4 4.6% 50 3.0% 90%(1-ε) 9M Impressions (100 ) 10k Impression CTR = 1.5% CTR = 5.1% CTR 4.74%
  42. 42. ε-Greedy MAB ε=0.10 45 10M Impressions 10%(ε) 1M Impressions( ) 1.1% 2.0% 8.2% 0.01% 4.6% 1.2% 5.2% 1.5% 0.2% 1.0% … 2.2% Best arm ( ) 3 8.2% 7 5.2% 4 4.6% 90%(1-ε) 9M Impressions (100 ) 10k Impression CTR = 1.5% CTR = 5.1% CTR 4.74% 10k Impression CTR Impressions CTR (3σ)
  43. 43. ε-Greedy MAB ε=0.10 46 10M Impressions 10%(ε) 1M Impressions( ) 1.1% 2.0% 8.2% 0.01% 4.6% 1.2% 5.2% 1.5% 0.2% 10 Impressions CTR Impressions CTR
  44. 44. ε-Greedy MAB ε=0.10 47 10M Impressions 10%(ε) 1M Impressions( ) 1.1% 2.0% 8.2% 0.01% 4.6% 1.2% 5.2% 1.5% 0.2% 1.0% … 2.2% Best arm ( ) 3 8.2% 7 5.2% 4 4.6% 90%(1-ε) 9M Impressions (100 ) 10k Impression CTR = 1.5% CTR = 5.1% CTR 4.74% CTR Impressions 99.7%(3σ)
  45. 45. ε-Greedy MAB ε=0.10 48 10M Impressions 10%(ε) 1M Impressions( ) 1.1% 2.0% 8.2% 0.01% 4.6% 1.2% 5.2% 1.5% 0.2% 1.0% … 2.2% Best arm ( ) 3 8.2% 7 5.2% 4 4.6% 50 3.0% 90%(1-ε) 9M Impressions (100 ) 10k Impression CTR = 1.5% CTR = 5.1% CTR 4.74% CTR Impressions 99.7%(3σ) CTR 3.0% 3.0% 3.0% 3.0% 3.0% 3.0%3.0%
  46. 46. ε-Greedy MAB ε=0.10 49 10M Impressions 10%(ε) 1.1% 2.0% 5.2% 1.5% 0.2% 1.0% … 2.2% Best arm ( ) 90%(1-ε) (100 ) 10k Impression CTR = 1.5% CTR = 5.1% CTR 4.74% CTR Impressions 99.7%(3σ) 3.0% 3.0%3.0% Optimal Arm Impressions (regret )
  47. 47. Thompson Sampling MAB ?
  48. 48. Thompson Sampling MAB • (arm) CTR Beta(a,b) . ( a=click, b=unclick ) 51 1 10% Impressions : 10 50 100 200 1k 10k 2 25% Impressions : 10 50 100 200 1k 10k
  49. 49. Thompson Sampling MAB • (arm) CTR Beta(a,b) ( a=click, b=unclick ) 52 1 10% ( ) CTR 15% 1 (10%<15%) 100 Impressions trial . Impression -> Impressions : 10 50 100 200 1k 10k
  50. 50. Thompson Sampling MAB • (arm) CTR Beta(a,b) ( a=click, b=unclick ) 53 2 25% 2 CTR 25%>15% ( ) . Impressions : 10 50 100 200 1k 10k
  51. 51. • 1K Impressions 54 
 CTR
  52. 52. • 10K Impressions 55 
 CTR
  53. 53. • 50K Impressions 56 
 CTR
  54. 54. • 100K Impressions 57 
 CTR
  55. 55. • 500K Impressions 58 
 CTR
  56. 56. • 1M Impressions 59 
 CTR
  57. 57. • 1M Impressions 60 
 CTR CTR (Arm)
  58. 58. • 1M Impressions 61 
 CTR CTR ( ) . .
  59. 59. • 1M Impressions 62 
 CTR CTR ( ) . . TS-MAB & Trade-off Regret( ) .
  60. 60. User Clustering • CTR . CTR : 7.6% A : 25% (30 ) B : 2.1% ( ) C : 7.1% ( ) CTR !63
  61. 61. User Clustering • X CTR • CTR 200 8,000
  62. 62. User Clustering • • 8 8 8,000 64,000 CTR
  63. 63. Clustering ? CB(image,Text) Feature User Feature [0.628, 0.88, 0.376, 0.065, 0.849] [0.508, 0.268, 0.193, 0.125, 0.425] [0.431, 0.077, 0.012, 0.07, 0.037] [0.915, 0.294, 0.713, 0.851, 0.423] [0.508, 0.268, 0.193, 0.125, 0.425] [0.607, 0.639, 0.554, 0.092, 0.297] [0.587, 0.319, 0.094, 0.173, 0.177] [0.409, 0.458, 0.48, 0.319, 0.783] [0.479, 0.434, 0.618, 0.297, 0.752] [0.467, 0.206, 0.905, 0.7, 0.568] , , , , , , , , !66 1 2 3 4 5 6
  64. 64. Clustering ? 14 CB(image,Text) Feature User Feature [0.628, 0.88, 0.376, 0.065, 0.849] [0.508, 0.268, 0.193, 0.125, 0.425] [0.431, 0.077, 0.012, 0.07, 0.037] [0.915, 0.294, 0.713, 0.851, 0.423] [0.508, 0.268, 0.193, 0.125, 0.425] [0.607, 0.639, 0.554, 0.092, 0.297] [0.587, 0.319, 0.094, 0.173, 0.177] [0.409, 0.458, 0.48, 0.319, 0.783] [0.479, 0.434, 0.618, 0.297, 0.752] [0.467, 0.206, 0.905, 0.7, 0.568] , , , , , , , , 8 (#0~#7) ?
  65. 65. - • • #1, #5, #6, #7 • #0, #2 #3
  66. 66. - • • #1, #5, #6, #7 • #0, #2 #3 /
  67. 67. #1 , ,
  68. 68. #3 , ,
  69. 69. / #3 , ,
  70. 70. Tag
  71. 71. Tag
  72. 72. / .
  73. 73. ? 77 #2 #1 #3 User Clustering Targeting CTR 1 : CTR 9.1% 2 : CTR 8.8% 3 : CTR 8.0% 4 : CTR 7.8% 5 : CTR 7.1% 6 : CTR 6.8% 7 : CTR 6.7% … MAB Ranking
  74. 74. ? 78 #2 #3 User Clustering Targeting Ranker MAB Ranking Targeter
  75. 75. + MAB
  76. 76. -ViewerEnd • : 
 CTR(%) 
 • • (Item Feature) • MAB(Multi Armed Bandit) !80
  77. 77. ? 81 #2 #3 User Clustering Targeting Ranker MAB Ranking Targeter
  78. 78. ? 82 (Item) Targeter Feature Targeting Ranker MAB Ranking
  79. 79. -ViewerEnd (CF) (Text) (Image) !83 1 2 3
  80. 80. - Item Features / Image (1) Image Feature , Text (2) Text Feature Feedback (3) CF-Feature !84
  81. 81. ? !85
  82. 82. ? !86
  83. 83. ? !87
  84. 84. 1 3 41 1 92 1 1 3 41 1 92 1 1 3 41 1 92 1 Image Text( ) CF( ) !88 1 3 41 1 92 1Image Style
  85. 85. Image Text( ) CF( ) Image Style Style transfer network !89 1 3 41 1 92 1
  86. 86. Image Style Text( ) CF( ) Image Object Detection Task Image Object Pre-trained VGG19 Model !90 1 3 41 1 92 1
  87. 87. Image Style Image CF( ) , Keyword Word2Vec !91 1 3 41 1 92 1
  88. 88. Image Style Image Text( ) CF( ) Matrix Factorization(ALS) with implicit feedback (Feedback) Item-User !92 1 3 41 1 92 1
  89. 89. !93 1 3 41 1 92 1 1 3 41 1 92 1 1 3 41 1 92 1 1 3 41 1 92 1
  90. 90. “ ” (%) !94
  91. 91. 0.4% 4.0% 2.9% 7.3% 2.7% 8.7% 6.7% 1.0% 1.9% 8.1% 0.4% 4.0% 2.9% 7.3% 2.7% 8.7% 6.7% 1.0% 1.9% 8.1% 0.4% 4.4% 2.9% 7.3% 2.3% 8.7% 0.2% 1.0% 1.9% 8.1% 0.4% 6.0% 2.9% 7.3% 2.7% 5.6% 6.7% 1.0% 1.9% 8.1% MAB (%) !95
  92. 92. 0.4% 4.0% 2.9% 7.3% 2.7% 8.7% 6.7% 1.0% 1.9% 8.1% 0.4% 4.0% 2.9% 7.3% 2.7% 8.7% 6.7% 1.0% 1.9% 8.1% 0.4% 4.4% 2.9% 7.3% 2.3% 8.7% 0.2% 1.0% 1.9% 8.1% 0.4% 6.0% 2.9% 7.3% 2.7% 5.6% 6.7% 1.0% 1.9% 8.1% !96
  93. 93. ? 97 (Item) Targeting CTR 1 : CTR 9.1% 2 : CTR 8.8% 3 : CTR 8.0% 4 : CTR 7.8% 5 : CTR 7.1% 6 : CTR 6.8% 7 : CTR 6.7% … MAB Ranking
  94. 94. + MAB !98
  95. 95. !99
  96. 96. 100 1 2 3 4 5 6 … 89 90 (Clicks) Impression
  97. 97. 101 1 2 3 4 5 6 … 89 90 (Use Coin) Impression
  98. 98. 102 User Cluster + + MAB MAB RankerTargeter = =
  99. 99. 103 + MAB Conditional Bandit Exponential Smoothing Seen Decay Soft User Clustering Retention Model Unbiased Most Popular Feature Matching Targeter = = Ranker
  100. 100. 104 • MAB
 - Bandit Algorithm = Thompson Sampling(
 - Reward = Click (with Unclick )
 - Play Arms = Cluster Most Popular 
 - None Stationary = Exponential Decaying • 2 
 - = # of clicks / # of impressions 
 - = # of use_coins / # of impressions 1. MAB
  101. 101. 105 • MAB
 - Bandit Algorithm = Thompson Sampling
 - Reward = Click (with Unclick )
 - Play Arms = Cluster Most Popular 
 - None Stationary = Exponential Decaying • 2 
 - = # of clicks / # of impressions 
 - = # of use_coins / # of impressions 1. MAB Use Coin( ) MAB Reward Use Coin, Click + User Coin by @brandon.lim
  102. 102. 106 1. MAB ? (Beta) (Alpha) -20% —> ? ? MAB ? -20%
  103. 103. 2. Conditional Bandit 107 1 2 3 4 5 6 … 89 90 by @troye.kwon
  104. 104. 2. Conditional Bandit 108 1 2 3 4 5 6 … 89 90 Impressions Reward=Click( ) α=click, β=unclick MAB by @troye.kwon
  105. 105. 2. Conditional Bandit 109 1 2 3 4 5 6 … 89 Impressions Reward=Click( ) α=click, β=unclick MAB Reward=Use Coin( ) α=use-coin, β=click MAB by @troye.kwon
  106. 106. 2. Conditional Bandit ? 110 (Beta) (alpha) ? ? - MAB .
  107. 107. 3. Retention Model • : . 
 , 
 “ ” . • • MAB 111 by @jinny.k + MAB Targeter Ranker
  108. 108. 3. Retention Model ? 112 by @jinny.k (CTR) (CVR) ? (CTR) ?
  109. 109. 4. Seen decay • : Negative Feedback • click impression Ranker • : (alpha) (Beta) 113
  110. 110. -> CTR 114 • • Hard Clustering(k-Means) —> Soft Clustering (pLSI) • Feature Matching • Targetting Genre/Tag Matching • MAB non-stationary Exponential Smoothing • Targeting Unbiased Most Popular • MAB Hyper parameter Turning
  111. 111. (%) !115 Soft Clustering (pLSI) Feature Matching Conditional Bandit Retention Model Exponential Smoothing MABUnbiased Most Popular
  112. 112. ? ? !116
  113. 113. !117
  114. 114. ? 118
  115. 115. ? 119
  116. 116. ? 120
  117. 117. ? 121
  118. 118. ? 122
  119. 119. 0.1% 123 = 3.96% = 6.70% = 2.63% = 0.07%
  120. 120. -> 4.56 124 = 3.96% = 6.70% = 2.63% = 0.07% 4.56
  121. 121. -> 4.56 125 = 3.96% = 6.70% = 2.63% = 0.07% 4.56 AB (<0.001) Feedback (>4.56days)
  122. 122. / , ?
  123. 123. . 127 Base : Editor’s ( X) 1.9% Alpha : 1 4.8% Beta : 2 5.5% Gamma : 3 6.5% CTR 1 2 3 4 …
  124. 124. . 128 Base : Editor’s ( X) 1.9% Alpha : 1 4.8% Beta : 2 5.5% Gamma : 3 6.5% CTR + 242% + 42% 1 2 3 4 … CTR
  125. 125. . ?
  126. 126. / ? ?
  127. 127. / ? -> ? ->
  128. 128. 133 YOU ?
  129. 129. . !134 |

×