a televÍziÓzÁs ÉlmÉnyÉnek nÖvelÉse ......source: d. zibriczky, z. petres, m. waszlavik and...

1

A TELEVÍZIÓZÁS ÉLMÉNYÉNEK NÖVELÉSE ALGORITMIKUS

MÓDSZEREKKEL, AVAGY PERSZONALIZÁLT TARTALOMAJÁNLÓ

SZOLGÁLTATÁS IPTV ÉS OTT RENDSZEREK SZÁMÁRA

Zibriczky Dávid, ImpressTV

2015-10-09

2

Who we are?

• Acquisition of IPTV, OTT and media business line of Gravity R&D (July 2014)

• Technical Centre in Budapest, sales and management from the UK

What we are doing?

• Providing personalized recommendations, data analytics, targeted advertising and audience measurement for corporate clients

• Main domains: Video-On-Demand, LinearTV, OTT, Advertisements

About ImpressTV

3

TV is dead?

4,000

40,000

1952 1962 1972 1982 1992 2002 2012

Radio Direct Mail Yellow Pages

Magazines Newspapers TOTAL TV

Source: B. Kitts - Tectonic Shifts in Television Advertising (RecSys 2014)

• TV and radio are not dead

• Temporary fall in 2008

• Newspapers, magazines and yellow

pages are in falling trend in 2000s

4

• More time spent watching media contents than ever

• Hundreds of channels, ten thousands of movies, millions of user uploaded videos

• Heterogeneous mixture of devices and contents

• Massive consumptional information

• Cloudization and centralization

Then and Now

5

Source: http://bigscreenglobal.com/

6

What is a Recommender System?

7

• For the consumers

› Content discovery (relevance, time, information filtering)

› Exploring new preferences (habits, engagement)

• For the business

› Improving KPIs, balancing consumption (long tail contents)

› Promotions, targeting, campaign, analitics, reporting

• For the vendors

› Data integration, insights, data science, optimization

› Technology challenges, deployment, maintenance

What does a Recommender System mean in TV business?

8

• Goals for consumers not necessarily equal to business!

• Simplified example 1: YouTube (User-generated contents)

› Goal of the business: Receive more income from advertisements

› Goal of the users: Having useful time by watching videos

› The goal is realized in the same way, more videos watched are good for both.

• Simplified example 2: Netflix (Video On Demand rental)

› Goal of the business: Increase the income of VOD

› Goal for the users: Watching interesting movies, spending less time for searching

› The goal is different. Netflix wants the users pay more. The users basically don’t want to spend more.

Difference between the goals

9

Challenges

10

Evolution of the Recommender Problem

Source: http://www.slideshare.net/xamat/recsys-2014-tutorial-the-recommender-problem-revisited

11

Time-dependency

1:0

0

1:3

0

2:0

0

2:3

0

3:0

0

3:3

0

4:0

0

4:3

0

5:0

0

5:3

0

6:0

0

6:3

0

7:0

0

7:3

0

8:0

0

8:3

0

9:0

0

9:3

0

10

:00

10

:30

11

:00

11

:30

12

:00

12

:30

13

:00

13

:30

14

:00

14

:30

15

:00

15

:30

16

:00

16

:30

17

:00

17

:30

18

:00

18

:30

19

:00

19

:30

20

:00

20

:30

21

:00

21

:30

22

:00

22

:30

23

:00

23

:30

0:0

0

0:3

0

Intraday Channel Popularity

12

Device-dependency

Source: B. Kitts - Tectonic Shifts in Television Advertising (RecSys 2014)

0.00

0.50

1.00

1.50

2.00

2.50

3.00

K2-11 T12-17 A18-24 A25-34 A35-44 A45-54 A55-64 A65+

Video on Mobile

Video on Mobile

0.00

0.50

1.00

1.50

2.00

2.50

K2-11 T12-17 A18-24 A25-34 A35-44 A45-54 A55-64 A65+

Video on internet

Video on internet

0.00

0.20

0.40

0.60

0.80

1.00

1.20

1.40

1.60

K2-11 T12-17 A18-24 A25-34 A35-44 A45-54 A55-64 A65+

Time-shifted TV

Time-shifted TV

0.00

0.20

0.40

0.60

0.80

1.00

1.20

1.40

1.60

K2-11 T12-17 A18-24 A25-34 A35-44 A45-54 A55-64 A65+

Traditional TV

Traditional TV

12-17 year olds

25-34 year olds

18-24 year olds

65+ year olds

13

Unification challenges

PVR

LiveTV

CatchUp

SVoD

TVoD

ADs

Devices

Social Networks

Metadata sources

TV contents (operators)

14

• Who is the front of the TV? (watching behavior pattern recognition)

• Cold-start problem, opt-out users (metadata enrichment, hybrid filtering)

• Noise filtering, lack of explicit data (filtering and weighting method, implicit feedbacks)

• Context-dependency (device, time, mood)

• Cross-domain recommendation, content centralization

• Time-dependent recommendable set (model-based methods)

• Popularity effect (diversification)

• Upselling, subscription (Live VOD)

• Offline vs. Online KPIs

Data Science Challenges

15

• Heterogeneous external data sources (data unification)

• Big data (scalable algorithms, distributed databases)

• Latency in data transfer, streaming service

• Responsivity, load, SLA (distributed systems, load balancer)

• Maintenance, service, follow-up

• Open source vs. self-development

• Distributed systems vs. single server

• Software-as-a-service vs. on-site deplyoment

Technology Challenges

16

Let’s do some math…

17

• Different preferences during the day

• Context: Time period

• Time period 1: 06:00-14:00

Tensor Factorization

R1 Item1 Item2 Item3 …

User1 1 …

User2 1 0 …

User3 …

…. … … … …

18


User1 1 …

User2 1 0 …

User3 …

…. … … … …


User1 0 1 …

User2 1 …

User3 1 …

…. … … … …




• Time period 2: 14:00-22:00

19


User1 1 …

User2 1 0 …

User3 …

…. … … … …


User1 0 1 …

User2 1 …

User3 1 …

…. … … … …


User1 1 …

User2 …

User3 1 1 …

…. … … … …




• Time period 3: 22:00-06:00

20


User1 1 …

User2 1 0 …

User3 …

…. … … … …


User1 0 1 …

User2 1 …

User3 1 …

…. … … … …


User1 …

User2 𝒓𝑢𝑖𝑡 …

User3 …

…. … … … …

QTq11 q21 q31 …

q12 q22 q32 …

P

p11 p12

p21 p22

p31 p32

… …

Tt11

t12

t21

t22

t31

t32

𝑹𝑵𝒙𝑴: preference matrix

𝑷𝑵𝒙𝑲: user feature matrix

𝑸𝑴𝒙𝑲: item feature matrix

𝑻𝑳𝒙𝑲: time feature matrix

𝑵: #users

𝑴: #items

𝑳: #time periods

𝑲: #features

𝒓𝒖𝒊t =

𝒌

𝒑𝒖𝒌𝒒𝒊𝒌𝒕𝒕𝒌

𝑹 = 𝑷°𝑸°𝑻Tensor Factorization

21

Hybrid Matrix

R1 Item1 Item2 Item3 …User

Meta 1

User

Meta 2…

User1 1 … 1 …

User2 1 0 … 1 …

User3 … 1 …

…. … … … … … … …

Item

Meta 11 1 … …

Item

Meta 21 1 … …

… … … … … … … …

22

Does the Recommender System work?

23

Case Study (VOD, Magyar Telekom)

-5%

0%

5%

10%

15%

20%

25%

30%

01

-01

01

-03

01

-05

01

-07

01

-09

01

-11

01

-13

01

-15

01

-17

01

-19

01

-21

01

-23

01

-25

01

-27

01

-29

01

-31

02

-02

02

-04

02

-06

02

-08

02

-10

02

-12

02

-14

02

-16

02

-18

02

-20

02

-22

02

-24

02

-26

02

-28

A/B testing VOD Revenue increase vs. External Baseline Group (7-day Moving Avg.)

ITV Algorithm 1 ITV Algorithm 2 ITV Algorithm 3

Changes in 3 Changes in 2

Changes in 2

Baseline 1

24

• Increasing trust in recommender system

• Contents selected via R4U are watched 40% longer and completed with almost twice more probability than in standard way.

• High correlation between Recall/MRR and Completed Waching Ratio

Source: D. Zibriczky, Z. Petres, M. Waszlavik and D. Tikk: EPG content recommendation in large scale: a case study on interactive TV platform. Machine Learning with MultimediaData – Special session at the 12th IEEE International Conference on Machine Learning and Applications (ICMLA'13), Miami, Florida, 2013.

Case Study (LinearTV, SaskTel)

0%

20%

40%

60%

80%

1 10 100

Average CTR vs. # of rec. requests from the

first use of RS

EPG/Zapp R4U Benefit

Watching LengthRatio

29,90% 42,02% +40,54%

Completed WatchingRatio

15,91% 30,53% +91,89%

25

• Heterogeneous data integration, cloud

• Content discovery and personalization over recommendation

• Productization over optimization

• Hybrid Filtering with Factorization Machines

• Content popularity prediction

• Mood detection, gamification

• Future TV: Personalized channels for each users, schedule recommendation

State of the art

26

Presented by:

Contact:

THANK YOU!

www.impresstv.com

David Zibriczky

Head of Data Science

[email protected]

a televÍziÓzÁs ÉlmÉnyÉnek nÖvelÉse ......source: d. zibriczky, z. petres, m. waszlavik and...

Documents