Transcript Chervet Benjamin

ACM IMC 2008
Watching Television
Over an IP Network
Meeyoung Cha
MPI-SWS
Pablo Rodriguez
Sue Moon
Jon Crowcroft
Xavier Amatriain
Telefonica Research
KAIST
U. of Cambridge
Telefonica Research
Presented and modified by : Chervet Benjamin
Part1.
IPTV overview
and dataset
Part2.
Analysis of
viewing patterns
Part3.
Channel change
probability
Internet TV (IPTV)
 Delivering television channels over an IP network
 20M subscribers worldwide in 2008
 Popular types
1. Telco’s nation-wide provisioned service
 By AT&T, France Telecom, Korea Telecom, Telefonica
2. Web TV
 Joost, Zatoo, VeohTV, Babelgum, BBC’s iPlayer
3. Box-based video-on-demand
 Apple TV, Vudu box, Sony’s Internet video link
Max Planck Institute
3
Why study TV viewing patterns?
 Understanding of human viewing behaviors
 Identify social and demographic aspects, user profiling
 Cost-efficient design of distribution architectures
 Evaluate existing designs and explore new ones
 Design better channel guides and advertisements
 Help people find interesting programs more quickly
Max Planck Institute
4
How me measure TV pattern today ?
 Nielsen TV rating
 Install a Box that register which Channel is watched.
 Every time an user watch the TV he must triggers a button.
 The data are then transferred and gathered
 All the household is monitored.
 Select representative samples
 Extrapolate statistics across a nation
Max Planck Institute
5
Challenges in traditional TV research
 Nielsen TV rating (con’t)
< Drawbacks >
 Potential bias in sampling
 Awareness to metering may alter user behaviors
 Only a few users willing to be so are monitored.
 Gathering data from a large number of samples
challenging
 IPTV allows for continuous and detailed TV analysis!
Max Planck Institute
6
A first study on Telco’s IPTV workloads
 Collected raw data of everybody watching TV
 A quarter million users from a large IPTV system
(entire subscribers within a nation)
 150 channels including various genres
(free-to-air, children, sports, movies, music, etc)
 Collected traces for 6 months
 Largest scale study on TV viewing patterns
 User base 10 times larger than the Nielsen’s
Max Planck Institute
7
Telco’s Network
Inter ISP
Optical Fiber
Router
Backbone
Copper Link
DSLAM
Backbone
Max Planck Institute
DSLAM
8
Telco’s IPTV service architecture
TV head end
customer premise
DSLAM
TV
All 150 channels
IP backbone
Max Planck Institute
1-2 channels
Set-top box
9
Data collection
 User’s channel change input
 IGMP messages collected
across all 700 DSLAMs
 Trace example





Timestamp
DSLAM IP
Set-top box IP
Multicast channel IP
Action (join or leave)
Max Planck Institute
Collected here
DSLAM
set-top-box
10
Part1.
IPTV overview
and dataset
Part2.
Analysis of
viewing patterns
Part3.
Channel change
probability
Channel holding times
 60% channel changes happen within 10 seconds
 Infrastructure must support fast channel changes
rough. 20 % of the channels
changes occurs between 1 min
and 1 hour of watching the
same channel
Max Planck Institute
12
Zipf distribution
Cumulative mass function
Probability mass function
The longer an user watch a channel, the less likely he
would be to change the channel in the next period.
Max Planck Institute
13
Assumptions about user modes
 Difficulty in inferring user away mode
 TV is OFF; or left ON without any viewer
 Determined active users as those who change
channels within a one hour threshold period
 Tested with longer thresholds
 Demarcate viewing from surfing by the minute
 Nielsen also uses 1 minute threshold
Max Planck Institute
14
Three user modes
 Each user in one of the three states at any given time
 Active session: consecutive time spent on surfing or
viewing
Max Planck Institute
15
Session characteristics
 Durations
 An average household watched 2.54 hours of TV and
6.3 channels (distinct) a day
 Each active session lasted 1.2 hours
 Each viewing event lasted 14.8 minutes
 Per content genre
 Average surfing time longer for documentaries and movies
(9-11 sec) than news, music, and sports (6-7 sec)
Max Planck Institute
16
Weekly pattern
Party Time
Sleep-in
Max Planck Institute
17
Diurnal pattern
 Viewing hours across users highly correlated
 Two peaks at lunch (3PM) and dinner (10PM) times
Max Planck Institute
18
Diurnal pattern with longer away threshold
 Applied 2-hour thresholds for certain genres
(movies, documentaries, sports, etc)
Max Planck Institute
19
Inferring user modes
 When static 2-hour threshold used for demarcating
active and inactive sessions
Max Planck Institute
20
Channel popularity
 90% of concurrent viewers watch 20% of channels
 Follow the Pareto principal (80% - 20% rules)
Max Planck Institute
21
Time evolution of channel popularity
 Viewer share of top channels higher at peak times
 Popularity of top channels reinforced at peak times
Max Planck Institute
22
Implications of viewing patterns
 60% of channel changes within 10 seconds (surfing)
=> Challenges for P2P-based IPTV systems
 User focus followed the Pareto principal
=> IP multicast not efficient for unpopular channels
Max Planck Institute
23
Part1.
IPTV overview
and dataset
Part2.
Analysis of
viewing patterns
Part3.
Channel change
probability
Channel change patterns
 Our goal is to understand
 How do people browse through channels?
Do they use electronic program guide?
 Do channel changes result in viewing?
 How do users join and leave a particular channel?
Max Planck Institute
25
Channel change probability
 Probability of joining channel y after joining channel x
60% linear
Max Planck Institute
26
Channel viewing probability
 Probability of viewing channel y after viewing channel x
67% non-linear
60% within genre
17% to the same
channel
Max Planck Institute
27
User arrival and departure rates
 Batch-like arrivals and departures
 Inheritance (continued viewing even after channel changes)
arrival
departure
Max Planck Institute
28
Implications of channel change patterns
 Disparity in how we change and view channels
=> Design of efficient program guide
 High churn (attrition rate), especially during
commercial breaks
=> Challenging for P2P-based IPTV systems
Max Planck Institute
29
Summary
 The first work to analyze television viewing patterns
from complete raw data of IPTV users
 Implications on the architecture
 Support fast channel changes
 Handle high churn during commercials
 Reflect Pareto channel popularity
 Implications on the viewing guide
 Devise a better way to browse channels
 Personalize suggestions for users
Max Planck Institute
30