Upload
ulysses-burdock
View
214
Download
0
Tags:
Embed Size (px)
Citation preview
Curing Discontent in Online Content AcquisitionNishanth SastryKing’s College London
Early use of mass media
http://www.watfordobserver.co.uk/nostalgia/memories/10099510.Coronation_treat_as_community_gathers_around_the_only_TV/
Picture from the TV broadcast of the Coronation of Elizabeth II in 1953, Watford
Today’s “TV” viewing
With Digital Media Convergence, TV is just another video app, accessed on-demand on the Web
What changed: Push PullSuperficially: audience to TV set ratio has decreased
At a fundamental level: audience per “broadcast” is lower “Broadcast” time is chosen by the consumer
Traditional mass media pushed content to consumer
Current dominant model has changed to pull
Generalizes to other mass media as well
Implications of the pull model Traditionally, “editors” decided what content got pushed when
Linear TV schedulers use complex analytics to decide “primetime”
Users get more choice with the pull model When to consume What to consume (from large catalogue)
Unpopular/niche interest content also gets a distribution channel, not just what editors decide to showcase/bless as “publishable”
Cheaper to stream over the Web to a single user than to broadcast (e.g. to operate/maintain equipment like high power TV transmitters) BUT: Cost of broadcast can be amortized across millions of consumers
Could be cheaper per user to broadcast than to stream
Research questions How does pull model impact delivery infrastructure?
Can additional load of on-demand pulls be reduced by reusing scheduled pushes?
How do users make use of flexibility afforded to them?
Were/are editors good at predicting popularity?
Is niche interest/unpopular content important to users?
How do users find unpopular content they like? Users help each other!
Understanding how and why users share their loves
Designing infrastructure to help users find most influential users for their topics of interest
WWW’13
ICWSM’12
ICWSM’13
ASE/IEEE Social Informatics’12
Data to answer the questions*
Nearly 6 million users of BBC iPlayer across the UK
32.6 million streams, >37K distinct content items
25% sample of BBC iPlayer access over 2 months
Five years of vimeo data (Feb’05 – Mar’10) Goes back to within 3 months of founding date
443K videos, 2.5 million likes, 200K users, 700K links
All content curation activity, Jan’13Pinterest (8.5 million users), Dec’12last.fm (nearly 300K users)
All tweets leading up to London Olympics (1.2 million), Closing Ceremony (~0.5 million), London Fashion Week (168K tweets)
WWW’13
ICWSM’12
ICWSM’13
ASE/IEEE Social Informatics’12
*Certain data can be made available upon request
Understanding and decreasing the network footprint of Catch-up TV
How does pull model impact delivery infrastructure?
Can additional load of on-demand pulls be reduced by reusing scheduled pushes?
How do users make use of flexibility afforded to them?
Were/are editors good at predicting popularity?
WWW’13
What users prefer to watch-I
• BBC proposes, consumer disposes!• Serials:~50% of content corpus; 80% of watched content!
Understanding and decreasing the Network Footprint of Catch-up TV-WWW’13
What users prefer to watch-II
Understanding and decreasing the Network Footprint of Catch-up TV-WWW’13
What users prefer to watch-III
Understanding and decreasing the Network Footprint of Catch-up TV-WWW’13
Abandoned
High preference for 30 and 60 min shows
Impact of pull on infrastructure
Understanding and decreasing the Network Footprint of Catch-up TV-WWW’13
On-demand spreads load over time
Linear TV schedulers seem to do a good job of predicting popularity!
On-demand more suited to web/pull than linear TV
• BUT: iPlayer traffic is close to 6% of UK peak traffic• Second only to YouTube in traffic footprint• Compare to adult video, a traditional heavy hitter. Most popular
adult video streaming sites have <0.2% traffic share
• BUT: amortized per-user, broadcast greener than streaming*
(using Baliga et al.’s energy model for the Internet)
*All channels except BBC Parliament, which has few viewers
Understanding and decreasing the Network Footprint of Catch-up TV-WWW’13
Still, can we decrease its footprint, please?
Yes, we can!
• DVRs have >50% penetration in US, UK• Many (e.g. YouView) don’t need cable• Could also use TV tuner and record on laptop
But, people don’t remember to record always
Understanding and decreasing the Network Footprint of Catch-up TV-WWW’13
Can we help users record what they want to watch?
Understanding and decreasing the Network Footprint of Catch-up TV-WWW’13
Speculative Content Offloading and Recording Engine
SCORE=predictor+optimiser
• Predict using user affinity for • Episodes of same programme• Favourite genres
• We can optimise for decreasing traffic or carbon footprint • Decreasing carbon decreases traffic, but not vice versa• Turns out we only take 5-15% hit by focusing on carbon
Understanding and decreasing the Network Footprint of Catch-up TV-WWW’13
Performance evaluation
• SCORE saves ~40-60% of savings achieved by oracle• Green optimisation saves 40% more energy at expense of 5% more traffic
Understanding and decreasing the Network Footprint of Catch-up TV-WWW’13
Compare SCORE relative to Oracle knowing future requests
Oracle saves:• Up to 97% of traffic• Up to 74% of energy
Not all of these savings come from predicting popular content
• Indiscriminately recording top n shows can lead to negative energy savings!
• Personalised approach necessary, despite popularity of “prime time” content
Understanding and decreasing the Network Footprint of Catch-up TV-WWW’13
How To Tell Head From Tail in User-generated Content Corpora
Is niche interest/unpopular content important to users?
How do users find unpopular content they like? Users help each other!
WWW’13
ICWSM’12
AAAIICWSM’12
The tail is heavy in users, not accesses
How to tell head from tail in User-generated Content Corpora- AAAI ICWSM’12
Like sets of many users are dense in tail items
How to tell head from tail in User-generated Content Corpora- AAAI ICWSM’12
Likers of tail content are geographically more diverse
How to tell head from tail in User-generated Content Corpora- AAAI ICWSM’12
Niche interest content rather than merely unpopular?
How do users find tail items?
How to tell head from tail in User-generated Content Corpora- AAAI ICWSM’12
Non-viral access predominates in popular items
Sharing the Loves: Understanding the how and why of online content curation
Understanding how and why users share their loves
Designing infrastructure to help users find most influential users for their topics of interest
WWW’13
ICWSM’12
ICWSM’13
Is niche interest/unpopular content important to users?How do users find unpopular content they like?
Users help each other!
AAAIICWSM’13
Sharing the Loves: Understanding the how and why of online content curation
Data reminder: All (38 million) Repins, (~20 million) Likes on
Pinterest Jan 13 All (90 million) Loves, (~60 million) Tags on last.fm
Dec 12
Survey respondents: 30 for Pinterest, 270 for last.fm
AAAIICWSM’13
Why people curate content
Sharing the loves: Understanding the how and why of online content curation- AAAI ICWSM’13
Curation comes up when search stops working – Clay Shirky
Last.fm
Curation: of personal or social value?
• Pinterest: (30 respondents, allow multiple answers)• 85% use it as a personal collection or scrapbook • 48% uses the site to display their content to others
• Last.fm: (279 respondents, allow multiple answers)• 39% tags tracks for personal classification• 39% tags to create a global classification (genres).
• The majority of respondents shared this view (last.fm):• “I find the social aspect more useful and interesting with
people I know, rather than developing new interactions based on music taste. ”
• BUT: one couple met on last.fm, started going to gigs together and are now happily married!!
Sharing the loves: Understanding the how and why of online content curation- AAAI ICWSM’13
Users mostly see it as personal effort, with exceptions
Despite unsynchronised personal effort, community synchronises on some topics!
Sharing the loves: Understanding the how and why of online content curation- AAAI ICWSM’13
Strong popularity skew, as in previous highlighting methods
Understanding how effective content curation happens
• Unstructured curation: Actions that simply highlight an item• e.g., love, like, ban, comment, shout
• Structured Curation: Actions that also organise item onto user-specific lists • e.g., pinning an item onto a user’s board, • attaching a user’s tag to a track
• Characteristics of effective curators: consistency, diversity…
Sharing the loves: Understanding the how and why of online content curation- AAAI ICWSM’13
Structured curation preferred for popularly curated items
Sharing the loves: Understanding the how and why of online content curation- AAAI ICWSM’13
How to curate: Consistent and regular updates attracts followers
Sharing the loves: Understanding the how and why of online content curation- AAAI ICWSM’13
The most important part of a curator’s job is to continually identify new content for their audience
-- Rohit Bhargava
How to curate: Diversity of interests attracts followers
Sharing the loves: Understanding the how and why of online content curation- AAAI ICWSM’13
IARank: Ranking Users on Twitter in Near Real-time, Based on their Information Amplification Potential Effective content curation is a highly demanding task
Consumers still need to find the best “editors” they want
Naturally self-limiting when it comes to high-volume events Olympics closing ceremony: 400K tweets in just over 3 hours
We can rank the most influential users e.g., PageRank
PageRank takes time to converge Ranks can change before!
IARank:ranks users by Information Amplification potential “Buzz” factor: how likely to be retweeted “structural advantage”: how good is your immediate
neighbourhood
Understanding how and why users share their loves
Designing infrastructure to help users find most influential users for their topics of interest
ICWSM’13
ASE/IEEE Social Informatics’12
Summary Characterising on-demand content consumption via 6
million users of BBC iPlayer
If broadcast is efficient, we should find ways to use it!
SCORE: personalised content offloading engine
Is niche interest/unpopular content important to users?
How do users find unpopular content they like? Users help each other!
Social curation complements search; effective curators are consistent and have diverse interests
Near-instantaneous reranking scheme for high volume content sharing systems like Twitter
WWW’13
ICWSM’12
ICWSM’13
ASE/IEEE Social Informatics’12
Curing Discontent in Online Content AcquisitionNishanth Sastry
King’s College London
http://www.inf.kcl.ac.uk/staff/nrs