Video Recommender in Viki (VikiでのVideoレコメンド事例)

Video Recommender in Viki

Takashi Umeda (梅田 卓志)

@umekoumeda

Co-work with Viki developers

2

Do you know Viki ?

3

Click!!

4

Free Video Streaming service

5

+ Anime, TV Drama & Movie

6

+ Subtitle

7

Existing recommender

Select videos randomly from videos

in same country & genre

Click Rate : 0.09%

8

Objective

Boost CTR of recommender in video page

Click Rate : 0.09%

9

Content attributes

User behavior

+

10

Content attributes

User behavior

+

11

Goal

Video

A

Video

B

Select videos for each video

: :

Video Video Video

Video Video Video

Recommend

Recommend

12

Recommendation by users’ behavior

Select videos watched by common users

TVNW: SBS

Genre: Romance

Year: 2013

TVNW: SBS

Genre: Action

Year: 2014

TVNW: TvN

Genre: Romance

Year: 2013

User 1

User 2

User 3

Watch WatchWatch Watch

Watch

video

Avideo

Bvideo

C

13

Recommendation by users’ behavior

TVNW: SBS

Genre: Romance

Year: 2013

TVNW: SBS

Genre: Action

Year: 2014

TVNW: TvN

Genre: Romance

Year: 2013

User 1

User 2

User 3

Watch WatchWatch Watch

Watch

video

Avideo

Bvideo

C

Select videos watched by common users

14

Recommendation by users’ behavior

TVNW: SBS

Genre: Romance

Year: 2013

TVNW: SBS

Genre: Action

Year: 2014

video

Avideo

B

Recommend “B” on page “A”

Recommend

15

But,...

16

Issue

Ratio of videos having results : 42%

Popular

VideosVideoVideo

Minor

VideosNo results

58%

42%

0%

20%

40%

60%

80%

100%

17

Content

attributes

User behavior

Episode / Parts

Other attributes

+

18

Content

attributes

User behavior

Episode / Parts

Other attributes

+

19

Parts

Merge parts into one video

Jungle

Emperor

Leo

Part1

130,976 videos 79,601 videos

Jungle

Emperor

Leo

Part2

Jungle

Emperor

Leo

20

Episodes

Merge episodes into one video

Doctor X

Episode 1

79,601 videos 22,844 videos

Doctor X

Episode 2Doctor X

21

Issue

Ratio of videos having results : 42%

Popular

VideosVideo Video

Minor

VideosNo results

75%

25%

75%

0%

20%

40%

60%

80%

100%

22

Issue

Popular

VideosVideo Video

Minor

VideosNo results

Remaining 25% How should we do ?

25%

75%

0%

20%

40%

60%

80%

100%

23

Content

attributes

User behavior

Episode / Parts

Other attributes

+

24

Procedure

Which attributes are same ?

TVNW: SBS

Genre: Romance

Year: 2013

TVNW: SBS

Genre: Action

Year: 2014

TVNW: TvN

Genre: Romance

Year: 2013

User 1

User 2

User 3

Watch WatchWatch Watch

Watch

video

Avideo

Bvideo

C

25

Probability of videos having same attributes

TV NW > Country > Genre > Actor, ..

TVNW: SBS

Genre: Romance

Year: 2013

TVNW: SBS

Genre: Action

Year: 2014

TVNW: TvN

Genre: Romance

Year: 2013

User 1

User 2

User 3

Watch WatchWatch Watch

Watch

video

Avideo

Bvideo

C

26

Application

Select videos focusing on TV NW

TVNW: SBS

Genre: Romance

Year: 2013

TVNW: SBS

Genre: Action

Year: 2014

video

Avideo

Bvideo

D

TVNW: SBS

Genre: Romance

Year: 2014

Rec.

Minor videos

(25%)

27

Application

Select videos focusing on TV NW

TVNW: SBS

Genre: Romance

Year: 2013

TVNW: SBS

Genre: Action

Year: 2014

video

Avideo

Bvideo

D

TVNW: SBS

Genre: Romance

Year: 2014

Rec.

Same TVNW’s videosMinor videos

(25%)

28

0.08

0.09

0.1

0.11

0.12

0.13

Old New

Click R

ate

[%

]

AB test

Click Rate : +32.4%

It uses AB test frame work ‘Turing’ developed by Ishan

29

DC

DC

DC DC

New recommender

It rolled out across the world

Front-end is developed by Huy & Yan Han

30

Click on the web !

31

Download Mobile App!

32

Appendix

33

Contents-based recommendation #1

Fix weights of attribute by user-behavior rec. result

Similarity(video A, video B) =

w1 * Genre Similarity (genre of video A, genre of video B)

w2 * Country Similarity (country of video A, country of video B)

w3 * Actor Similarity (actors in video A, actors in video B)

:

Weights are fixed by user-behavior rec. result

• If attributes are matched, it’s 1.

• Otherwise, it’s 0.

34

Contents-based recommendation #2

• Fix weigh by using user behavior recommeder result

• Estimate similarity for videos which have no results

1.Training 2.Test

Video

A

Video

B

Genre

A

Genre

B

Jaccard

similarity

1v 2v kpop rock 99.0

1v 3v kpop jazz 3.1

1v 4v kpop classic 2.1

Video

A

Video

B

Genre

A

Genre

B

Jaccard

similarity

5v 2v jpop rock ?

5v 3v jpop jazz ?

5v 4v kpop classic ?

Similarity (videoA, videoB ; W)

Fix weight

Similarity (videoA, videoB ; W)

Estimate jaccard similarity