1 developing a predictive model of quality of experience for internet video athula balachandran...

Post on 14-Dec-2015

215 Views

Category:

Documents

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

1

Developing a Predictive Model ofQuality of Experience for Internet Video

Athula Balachandran

Carnegie Mellon University

2

Content

1 Why do quality of experience(QoE)?

2 Where are the challenges?

3 How to do ?

4 Implication and evaluation

3

Content

11 Why do quality of experience(QoE)? Why do quality of experience(QoE)?

2 Where are the challenges?

3 How to do ?

4 Implication and evaluation

4

Why?

Tow main revenue models:

1.Subscription

2.Advertisement

advertisement

subscription

5

Why?

Tow main revenue models:

1.Subscription

2.Advertisement

advertisement

subscription

The more you watch, The more we profit.

6

Why?

Tow main revenue models:

1.Subscription

2.Advertisement

advertisement

subscription

The more you watch, The more we profit.

Improving users’ quality of experience(QoE) is crucial

7

Content

1 Why do quality of experience(QoE)?

22 Where are the challenges? Where are the challenges?

3 How to do ?

4 Implication and evaluation

8

Era changed & Requirement improve

■Video quality: PSNR(Peak Signal-to-Noise Ration)

<average bitrate, join time, buffering ration, rate of buffering>

■ User experience: User Opinion Scores

User’s Engagement-centric( viewing time , number of visits)

9

Era changed & Requirement improve

■Video quality: PSNR(Peak Signal-to-Noise Ration)

<average bitrate, join time, buffering ration, rate of buffering>

■ User experience: User Opinion Scores

User’s Engagement-centric( viewing time , number of visits)

Average bitrate: HD(High-Definition) SD(Standard-Definition) LD(Low-

Definition)

Join time: load time

Buffering ratio: buffer_time/(buffer_time+play_time)

Rate of buffering: frequency of buffering

10

Challenge scope

Video quality User engagement

1.Video quality interdependence

2.Complex relationship

3.Confound factors influence

bitrate

Join time

bufratio

Visits num

Viewing time

Time of day

Type of video

11

Challenge 1 quality interdependence

Among video quality are subtle interdependence

1.Video quality interdependence

bitrate

Join time

bufratio

12

Challenge 2 complex relationship

Relationship between quality and engagement2.Complex relationship

bitrate

Join time

Visits num

Viewing time

13

Challenge 3 confound factors

Confound factors affect quality -> engagement

3.Confound factors influence

Type of Device

Type of video

14

Content

1 Why do quality of experience(QoE)?

2 Where are the challenges?

33 How to do ? How to do ?

4 Implication and evaluation

15

Compare current work

16

Compare current work

1.Model consider complex relationship and confound factors

2.Provide strategy for system design

17

Requirements for predictive model

Tackling relationship (quality->engagement) and interdependency (among quality)

Tackling confounding factors 1.Identifying the import confounding factors2.Address the confounding factors

18

Compare methods for tackling relation

Compare the accuracy of tackling relationship( quality -> engagement) and interdependency (among quality)

19

Confounding factors-Identify

----3 round filter for all possible Con. Factors----

20

Confounding factors-Identify

----3 round filter for all possible Con. Factors----

Round1: calculate Information Gain

Entropy:

Condition entropy:

Information gain:

Relative Information gain:

21

Confounding factors-Identify

----3 round filter for all possible Con. Factors----

Round1: calculate Information Gain

Entropy:

Condition entropy:

Information gain:

Relative Information gain:

……

22

Confounding factors-Identify

----3 round filter for all possible Con. Factors----

Round1: calculate Information Gain

Round2: compare Compacted Decision Tree

23

Round2: compare Compacted Decision Tree

24

Round2: compare Compacted Decision Tree

GE-1: A1<=4 ■ [sup=40%,con=100%]GE-2:A1>4 ○ [sup=60%,con=82%] Except: A1>7,A2<=2.5 ■ [sup=20%,con=86%] Except:A1<8.5,A2>1.8,A2<=2.5 ○ [sup=14%,con=100%]

25

Round2: compare Compacted Decision Tree

GE-1: A1<=4 ■ [sup=40%,con=100%]GE-2:A1>4 ○ [sup=60%,con=82%] Except: A1>7,A2<=2.5 ■ [sup=20%,con=86%] Except:A1<8.5,A2>1.8,A2<=2.5 ○ [sup=14%,con=100%]

■ ○ Total

A1<=4 24 0 24

A1 > 4 6.45 29.55 36

Total 30.45 29.55 60

χ2(A1<=4 ■) = (24-30.45)^2/30.45 = 1.37χ2(A1>4 ○) = (29.55-29.55)^2/29.55 = 0

26

Round2: compare Compacted Decision Tree

GE-1: A1<=4 ■ [sup=40%,con=100%]GE-2:A1>4 ○ [sup=60%,con=82%] Except: A1>7,A2<=2.5 ■ [sup=20%,con=86%] Except:A1<8.5,A2>1.8,A2<=2.5 ○ [sup=14%,con=100%]

■ ○ Total

A1>4,A2<=2.5,A1>7 6.45 1.05 7.5

A1>4,A2<=2.5,A1<=7 0 7.5 7.5

Total 6.45 8.55 15

χ2(..A1>7 ■) = (6.45-6.45)^2/6.45 = 0χ2(..A1<=7 ○) = (7.05-8.55)^2/8.55 = 0.26

1.Dif 2.sig 3. n(current) >= n(former) +1

27

Round2: compare Compacted Decision Tree

GE-1: A1<=4 ■ [sup=40%,con=100%]GE-2:A1>4 ○ [sup=60%,con=82%] Except: A1>7,A2<=2.5 ■ [sup=20%,con=86%] Except:A1<8.5,A2>1.8,A2<=2.5 ○ [sup=14%,con=100%]

28

Confounding factors-Identify

----3 round filter for all possible Con. Factors----

Round1: calculate Information Gain

Round2: compare Compacted Decision Tree

29

Confounding factors-Identify

----3 round filter for all possible Con. Factors----

Round1: calculate Information Gain

Round2: compare Compacted Decision Tree

Round3: compare Tolerance

30

Confounding factors-Identify

----3 round filter for all possible Con. Factors----

Round1: calculate Information Gain

Round2: compare Compacted Decision Tree

Round3: compare Tolerance

31

Confounding factors-Identify

----3 round filter for all possible Con. Factors----

Round1: calculate Information Gain

Round2: compare Compacted Decision Tree

Round3: compare Tolerance

32

Confounding factors-Address

Compare tow candidate way:

33

Confounding factors-Address

Compare tow candidate way:

Add as a new feature

Split data by Con. factors

34

Content

1 Why do quality of experience(QoE)?

2 Where are the challenges?

3 How to do ?

44 Implication and evaluation Implication and evaluation

35

Implication for system design

For an example model:buffering ratio, rate of buffering,

join time

Estimate all possible combinations

36

Implication for system design

Thanks!

top related