Social Activity as the New TV Guide

Page created by Brandon Hamilton
 
CONTINUE READING
Social Activity as the New TV Guide
Social Activity as the New TV Guide
                                Keith Mitchell1/2, Nicholas Race1/2, Adam Lindsay2
 1                                                    2
     21media innovations ltd, InfoLab21, Lancaster, UK; School of Computing and Communications, Lancaster University, Lancaster, UK

              E-mail: keith@21media.tv, race@comp.lancs.ac.uk, atl@comp.lancs.ac.uk

Abstract: The huge catalogues of video items currently               nation during the televised Leaders’ Debate. Our recent
available through live, on-demand and catchup TV                     research is also exploring this direction and specifically
based services coupled with the ever expanding range                 looks to providing awareness of friend activity within a
of online video services (Vimeo, YouTube, Netflix)                   social network and exposing this on a TV service user
provides users with unprecedented levels of choice.                  interface. Ultimately, we seek to help the user answer the
Not only is the choice of content expanding but so too               key question “What should I watch now”? to achieve this
are the range of formats, delivery mechanisms and the                we expose information relating which helps to answer
viewing devices (PC, TV/STB, mobile device). In order                additional questions viewers have which are “what live
to navigate these repositories sensibly recommender                  content is available now?”, “is there anything I may have
systems provide a key role in helping a user filter                  missed?” and “what are my friends watching now or what
items of interest. In this paper, we describe our initial            have the watched recently?”.
work on integrating an IPTV service with social                      The remainder of this paper describes the social IPTV
networks in order to support the personalisation                     system developed and which has been operational at the
process by exploiting the social graph.                              University of Lancaster, UK for 9 months. We outline
                                                                     some initial results and conclude by defining a number of
Keywords: Social TV, IPTV, Recommender Systems,                      key areas of future work which need to be addressed by
Social Graph, Content Distribution                                   the research and commercial communities.

                  1     INTRODUCTION                                                2     RELATED WORK
                                                                     Traditionally, recommender systems (RS) track user
The Internet is increasingly being used to distribute both
                                                                     activity and suggest suitable alternatives to each
real-time and on-demand high bandwidth multimedia
                                                                     individual viewer based on their previous activity. A
content to large audiences. IPTV specifically, through
                                                                     number of previous (and existing) systems that
services like BBC iPlayer, Hulu, TV Catchup, is on the
                                                                     recommend TV shows or movies using this technique
increase and traditional ‘over the air’ broadcasters are
                                                                     include, Netflix and LoveFilm which are both online TV
now looking to new delivery mechanisms via the Internet.
                                                                     provider and DVD rental services, WeOnTV [1]. Activity
These new delivery mechanisms offer flexible viewing
                                                                     in these contexts is normally based on explicit user
for the end user. Users are increasingly able to watch
                                                                     activity, such as rating, voting, commenting. However, it
what they want, when they want and within IPTV systems
                                                                     is well understood that the majority of viewers are
where content is increasingly being delivered to a browser,
                                                                     consumers of content and not necessarily contributors, so
users are also able to view related web content
                                                                     one may argue that it is not reliable to exploit these
simultaneously. In these scenarios where the linear
                                                                     factors alone.
broadcast model increasingly gives way to on-demand
access, there is currently a limited understanding of the            As a brief overview, in general, recommender algorithms
emerging usage patterns and context of use.                          recommend items similar to the ones the user liked in the
                                                                     past and can be classified into two families:
We believe that a better understanding of the consumption
patterns are required in order for recommendation                    • content-based filtering CBF) finds similarity between
systems to accurately recommend suitable content to end-                  items on the basis of the items content (e.g., gender,
users. Additionally, we believe that traditional content-                 director, actors, plot)
based filtering and collaborative filtering techniques               • collaborative filtering (CF) finds similarity between
would benefit from the use of activity information                        items on the basis of collaborative information about
gleaned from social networks, which essentially represent                 users, i.e., they use the opinions (known as ratings)
the human network. It is well understood and                              expressed by the community of ITV users.
acknowledged that TV viewers like to discuss the shows               The most popular recommender systems are based on
they watch, so called water cooler conversations. A                  collaborative filtering, the biggest advantage over
number of system have been developed which encourage                 content-based systems being that collaborative filtering
real-time discussion and interaction and in the United               relies only on explicit or implicit opinions expressed by
Kingdom, the 2010 Election led to Broadcasters                       users while explicit content description are not required.
experimenting with real-time video delivery on the web
                                                                     However, in order to accurately provide recommendations,
which also showed Twitter and Facebook feeds so that
                                                                     collaborative systems must derive users’ preferences
viewers could comment and observe the mood of the

Keith Mitchell, 21media innovations ltd, KBC, InfoLab21, Lancaster, LA1 4WA, +44 (0)161 498 5860, keith@21media.tv
toward items, which is done through analysis of viewers’        operational since November 2009 at the University of
past interaction with the TV programmes. The RS must            Lancaster, UK with over 6,500 students on campus.
thus track what the user are watching, in other words it
must collect viewers’ ratings. However, collaborative           3.1 Hardware Architecture
filtering requires large numbers of ratings before they can     The backend hardware includes three IPTV gateway units
make reasonable suggestions back to the community and           (also referred to as a Head End), a single Network
these systems additionally suffer from ‘cold-start’             Attached Storage (NAS) drive, a web server and a
problem which refers to situations where there are only         database server. To provide a live channel stream, digital
few ratings to base recommendations on.                         terrestrial (DVB-T) and free-to-air satellite (DVB-S) are
Our initial work in this area is looking into implicit          taken off-air and fed directly to each of the three IPTV
‘interest profile generation’ which includes data such as       gateway units. Each gateway then multicasts individual
linger and viewing times, channels changes/requests, and        MPEG-2/H.264 encoded channels over the University’s
in fact any interaction with the TV based service. The          private core and residences network. To serve the web-
focus of our work is on understanding the user-journey in       based client an Apache HTTP server with PHP is
detail and moving away from focusing on a specific              employed to provide a mixture of dynamic AJAX-enabled
content type (live or on-demand), genre or specific event,      HTML pages and XML feeds in order to build the client
in order to develop more generalisable solutions. A             and its supporting pages, as described in [7].
distinguishing element of the work presented here is the        3.2 Client Interface Components
development of our own measurement platform which
accurately tracks the user journey (per click or event)         The web based interface is shown in Figure 1 and consists
irrespective of the context of use, e.g. VoD or live streams.   of several components including the primary navigation
A further novel aspect of our approach is to monitor and        menu enabling a user to choose between live TV, radio
investigate the impact of the social graph (gleaned from        and VoD content, the video window based on the VLC
social networks such as Facebook) and to understand its         media player plug-in, the content carousel, filters for
potential influence on patterns of access.                      reordering the content carousel and the social panel, for
                                                                showing information retrieved from social networks.
This contrasts previous work which tends to rely either on
passively collected traces, publicly available traces of
static content (e.g. Internet Traffic Archive [5]), or
privately obtained (anonymised) logs of more dynamic
streaming content from larger networks such as Akamai’s
[11]. In Hei et al [6] for example, the performance of
PPLive using traces collected through passive packet
sniffing is studied. In [10], Silverstone et al in use a
similar approach to compare PPLive, PPStream, SopCast,
and TVAnts data collected during the FIFA World Cup in
2006. In general, the aforementioned works could be
regarded as providing “black-box or out-of-band”
measurement analysis as they do not instrument the end-                 Figure 1: Wireframe highlighting UI design
user client application itself.
                                                                Figure 2 shows the real web interface with EPG data
Within the broadcast Television domain, Social TV               being shown to the user along with a live video feed and
currently refers to the ability to allow people                 tweets related to the video content obtained via Twitter.
geographically separated but watching the same program          From here, a user may hover their mouse over a
to feel as if they were co-located and having a shared          programme title to view a synopsis of the programme or
experience [1]. These systems tend to be before the mass        hover over a channel logo to display channel-specific
popularity of social networks such as Facebook, Bebo and        Facebook-related social data, namely who else is
Twitter. Social TV increasingly refers to the integration       watching right now or watched recently.
of TV services with social networks such as Channel 4’s
Test Tube Telly allowing users to broadcast the
programme they are watching to Facebook and Twitter.
IPTV services such as Hulu and Zatoo provide some basic
social features in that they offer direct communications
channels between viewers [1]. However, as yet, none of
these leverage the social graph to support the knowledge
acquisition and decision making processes.
                  3     SOCIAL TV
This section describes briefly the hardware and software
architecture related the social TV system (known as
ResNet.TV) we have developed and which has been
Figure 2: The main client viewing page                 developing a flexible context model that may be
The personalisation of content is achieved via filters            applicable to a wide range of viewing contexts. Our
located on the right of the interface and allows users to         model builds on our own previous work on applications
reorder the results by:                                           such as the GUIDE context-aware tourist guide [3] and is
                                                                  based upon Dix et al’s [4] taxonomy for context with
• Popularity: users may choose between channel order,             respect to human-computer interaction of a system. This
     channels that are popular at that moment, or                 broad taxonomy can be applied to encapsulate the
     channels which have been popular that day.                   particular context types relevant to, and focussed on, in
• Social Awareness: allows users to restrict the usage            our work, which are the who (user identity), when (date
     data displayed to them which calculates popularity to        and time), where (user location), what (activities and
     relate to everyone or just their Facebook friends.           history of activity) and how (access technology/device)
In addition to the core navigation and content retrieval          and, uniquely, social (people in the virtual or physical
features, we additionally provide access to third party           vicinity) and resource management (device or network
services such as YouTube, Twitter and Facebook. The               resource usage or availability).
YouTube widget shows context-sensitive popular videos             Context data pertaining to user identity and activity (who
based on the currently playing programme’s title. The             and what) is derived implicitly by the system during
social panel offers either a contextual Twitter feed              interaction with the system. In general, user profiles may
produced by querying using the currently playing                  be constructed manually (explicitly) by the user or
programme’s title, or a Facebook feed of which                    automatically (implicitly) by a system. We have chosen
programmes each friend is currently watching or                   to focus on implicit profile creation first based on each
programme they watched last.                                      user’s content usage history since this provides the data
                                                                  necessary to provide the required intelligence.
3.3 Monitoring Platform                                           Additionally, from our previous work involving explicit
JavaScript is used extensively to instrument the user             user profiling [3] we know that is often difficult to rely on
interface in order to gather activity information. Each user      this model alone [4].
interface request generates an AJAX HTTP POST request
to the web server, which in turn is processed by a PHP                                 4     RESULTS
script, generating an entry in the central database hosted        In our recent work [7,8] we have provided comprehensive
on the database server. The POST request contains                 results pertaining to the analysis of our initial trials
several parameters and some are sent with every request           covering a 7 month period from October 2009 to May
(highlighted using italics), whereas other parameters are         2010. Throughout this period we have recorded over 4
specific to the type of event generated.                          million discrete events, corresponding to over 4,500
• General Events: Date/time of event, User IP address,            unique users, over 65,00 browser sessions and
      Browser session ID, Interface version (e.g. staff or        approximately 20% of active Facebook accounts. In
      student), Time between last channel change, Browser         summary, during this period we have found that:
      or tab focused/lost focus.                                  • Significant correlations exist between individual user
• Media Playback Events: Channel currently being                       and the overall (i.e. global) popularity of programmes
      watched, Channel launch / channel exit, play/pause,              on offer and that these correlations could be
      Full-sceen or window mode, volume up/down.                       leveraged in order to predict the potential popularity
• Navigation Based Events: Main navigation bar                         of a particular piece of content given a user’s content
      selection, Channel order/by popularity filtering                 consumption history.
      selection, Social filtering selection, Category selected.   • To predict global popularity of a piece of content
• Social Interaction: YouTube video selected, Twitter                  some of the most telling measures we found involved
      post hyperlink/profile hyperlink followed, Facebook              a user’s historical average percentage watches (i.e.
      login/logout, Facebook unique ID, Thumbs up/down.                users whom tend to watch all of a given programme),
                                                                       number of (Facebook) friends also using the service
Upon receiving the POST request, the PHP script                        and details relating to the way in which they search
calculates the position of the currently playing                       and navigate (in our system this relates to carousel
programme in the popular right now/my (Facebook)                       reconfigurations).
friends carousel, regardless of the actual carousel
configuration selected by the user. So, for users that            • Almost 25% of active sessions include explicit
always watched the same programme as their Facebook                    Facebook logins while using the service. Although
friends, their average calculated ‘position in friends list’           this is not a particularly high figure at this early stage,
metric would be 1, representing that the programmes they               we currently do not include any guidance on the
watch are always in the first position in the carousel.                benefits of using the ‘Connect with Facebook’ feature
                                                                       or describe how it works, what it does or what the
It is this instrumentation mechanism which provides us                 privacy implications are. Our landing page simply
with very detailed records of each user’s session. These               states ‘Connect to your Facebook account to see what
records are kept for the purpose of presenting content                 your friends are currently watching.’ While this was
filtered by popularity as well as for historical analysis.             an intentional decision so that we could measure the
For the purposes of our research, we are interested in
willingness of users to sign-up without being made        aims to highlight some of these areas as well as highlight
     aware of the implications, we were surprised by this      some of the lessons we have learned.
     high figure overall given the very strong references      The overriding challenge we feel is worth keeping in
     and concerns related to privacy expressed during our      mind is ‘scale’ and especially ‘scale to the web’ in order
     focus groups.                                             that any approach be effective and/or adopted given the
• By comparing usage of our system between those               nature of future TV services being based on a hybrid of
     users that signed in with Facebook and those which        traditional broadcast and more recent web based standards
     chose not too, we found an interesting result. The        and technologies but which, additionally, need to address
     more often a ‘social user’ visits the service and         the following.
     experiences the content the more unique channels
     they tend to experience. In contrast, those that          5.1 Findability
     choose not to use the social features experience fewer    Findability (not content) is King! Namely, the ability for
     channels. While this requires further investigation       viewers to efficiently and effortlessly discover new (and
     we speculate that this is due to the social awareness     existing) content. This simple process would benefit from
     features built in to the interface which exposes what     advances in multimedia search and discovery mechanisms
     other friends are or have been watching and that this     which can intelligently leverage content, content metadata
     does impact choice.                                       and user generated metadata, personalisation and accurate
• Difference were present between the number of                recommendations. A goal could be to support retrieval of
     different channels watched and how many of these          content based on questions such as “Show me all the
     were in the top 10 most popular each day. We found        content from this week which my friends have watched
     that the users which made use of the social awareness     which I have not and which I am interested in”.
     features was higher and that they watched 4/10            5.2 Visualisation
     compared to 2/10 on average.
                                                               Search and discovery of new content requires adequate
• The average length of time the first channel selected
                                                               representation to the end-user. In systems which are
     was viewed is increasing over time, which suggests
                                                               aggregating live, on-demand/catch-up and web content,
     that users are choosing a specific channel
                                                               new approaches to the visualisation and presentation of
     immediately rather than browsing (or channel
                                                               EPG data beyond the traditional ‘grid’ are required.
     hopping) on arrival at the service.
                                                               Mechanisms for appropriately displaying and enabling
While the above provides a brief summary of our findings,      navigation of content is, as has been discovered in the EU
they also pose many questions or lines of enquiry which        funded P2P-Next project, constrained by the end
are open to investigation still. We believe there are          system/device which may be a pc/laptop, mobile or low
lessons to be learned from studying of usage such as a         powered CE device. The ability to re-use and/or re-
user’s actual viewing behaviour, their social connections      purpose data and customise for individual end-user
and interactions, and their high level interactions with the   context is a major challenge especially when one wishes
system as a whole (i.e. interface events). The following       to offer the same or similar user experience.
section defines a series of discussion topics and areas for
future work and potential collaboration.                       5.3 Data Aggregation
                                                               Data aggregation requires new methods for supporting the
                                                               exponential growth and reuse of new content, unified and
         5     FUTURE CHALLENGES                               simple access to distributed multimedia content assets of
In this paper we have thus far summarised our approach         diverse formats, efficient methods for extracting data
and high level results relating to the development of a        from activity streams and the social graph (e.g. activity,
Social IPTV platform. Our ongoing and future work for          status, likes, comments, ratings, etc) which exploit current
the development of the Social TV platform is based on          standards and inform new standards in television content
the forwarding of two key areas:                               metadata enrichment.
     • User Interface to TV systems: Namely the                5.4 Metadata
         advancement of applications which enable the
         personalisation and recommendation                    Use of linked data, professional and user generated
                                                               semantic metadata specification and encouragement
     • Distribution mechanism and effective delivery           participation of tagging and adding new metadata to better
         of media given the heterogeneous nature of            support filtering/personalisation and the avoidance of the
         devices and communications networks and               ‘cold start’ problem. Mechanisms for representing user
         technologies                                          data statistics, strength of user interest(s) and user
Through the design, development and initial evaluation of      context(s) in machine-readable format, e.g., RDF, FOAF.
our system with the student population we have identified
a number of areas of future work or themes which we feel       5.5 Identity and Privacy
are still inadequately addressed by the network and            Personal identity is core to any social based network or
electronic media (NEM) communities. This last section          service. While identity and identity management are facts
                                                               of life for web based services, this is still a new
phenomena for TV, especially the notion of logging on to        temporal indexing into their content, then they tap into the
your TV. However, the merging of the web and TV                 so-called wisdom of the crowd to index (and implicitly
worlds inevitably means identity management will                rate) interesting fragments of their content.
increasing becoming part of TV and especially IPTV
based offerings. This area of research and development                            6      CONCLUSIONS
raises questions surrounding new identity services,             This paper has summarised our recent work on the
services and/or federated access between services and           development of an IPTV service which exploits the social
service providers, single sign-on. Further, clear and           graph for personalisation. Through the initial evaluation
unambiguous controls of user identity management and            of this system we have obtained some valuable lessons
profile managements are required.               Specifically,   and results which highlight the potential effectiveness of
unobtrusive methods for making clear the privacy                exploiting user activity information for both end-user
implications of changes to profiles and user settings.          (personalisation) and (intelligent) content distribution
Additional features to enable a user to maintain control of     systems. Our paper has concluded by introducing a
their profile(s), virtual identities, content and contextual    number of broad themes and a series of open questions of
metadata and mechanisms to ensure system wide privacy           thoughts which act as our motivation for continued and
integrity which are complicated when aggregating data           new research projects within the Network and Electronic
from various repositories is introduced? Major questions        Media (NEM) domain.
are raised when one considers how one ensure you are not
unwittingly revealing sensitive information during the
aggregation and enhancement process?                            References
                                                                [1] Abreu, J. F., Almeida, P., Pinto, R., and Nobre, V. 2009.
5.6 Content                                                     Implementation of social features over regular IPTV stb. In Proceedings
Licensing and copyright issues will remain and debates          of the Seventh European Conference on European interactive Television
                                                                Conference (Leuven, Belgium, June 03 - 05, 2009). EuroITV '09. ACM,
over the benefit of closed, proprietary systems versus          New York, NY, 29-32.
open systems will continue. Openness, specifically, open        [2] S. Ali, A. Mathur, and H. Zhang. Measurement of Commercial
data or linked data is important for tomorrow’s media           Peer-to-Peer Live Video Streaming. In Proceedings of International
distribution systems so that media assets can be ingest         Workshop on Recent Advances in Peer-to-Peer Streaming, 2006.
                                                                [3] Cheverst, K., Mitchell, K., Davies, N., and Smith, G. 2000.
into the network, accurately tracked (found), stored            Exploiting context to support social awareness and social navigation.
efficiently (in data centres, the cloud, community clouds,      SIGGROUP Bull. 21, 3 (Dec. 2000), 43-48.
peers) and distributed.                                         [4] Dix, A., T. Rodden, N. Davies, J. Trevor, A. Friday, K. Palfreyman
                                                                (2000), Exploiting space and location as a design framework for
5.7 Distribution                                                interactive mobile systems, ACM
                                                                [5] P. Danzig, J. Mogul, V. Paxson, and M. Schwartz. Internet Traffic
Distribution, efficient distribution on web scale is vital      Archive. http://ita.ee.lbl.gov.
and requires mechanisms, potentially adaptive or hybrid         [6] X. Hei, C. Liang, J. Liang, Y. Liu, and K. W. Ross. A
which can leverage / P2P/ CDN and other IP based                Measurement Study of a Large-Scale P2P IPTV System. IEEE
                                                                Transactions on Multimedia,9(8):1672–1687,Dec,07
technologies. Metadata distribution as well as the content      [7] Jones, A., Mitchell, K., and Race, N. J. 2010. TriggerTV:
distribution itself is also critical. Questions related to      exploiting social user journeys within an interactive TV system. In
how best to capture a user’s context automatically (i.e.,       Proceedings of the 20th international Workshop on Network and
minimise the explicit effort required by a user) as well as     Operating Systems Support For Digital Audio and Video (Amsterdam,
                                                                The Netherlands, June 02 - 04, 2010). NOSSDAV '10. ACM, New York,
how best to interpret and act upon this data to support         NY, 51-56.
efficient delivery are raised.                                  [8] K., Mitchell, A., Jones, J., Ishmael, N., Race, Social TV: Toward
                                                                Content Navigation using Social Awareness. In proceedings of
5.8 Micro-assets                                                EuroITV’10, 9th-11th June, Finland.
                                                                [9] Qiu, T., Ge, Z., Lee, S., Wang, J., Xu, J., and Zhao, Q. 2009.
By using fine-grained temporal indexing of viewing times        Modeling user activities in a large IPTV system. In Proceedings of the
(which are already available), broadcasters can get near-       9th ACM SIGCOMM Conference on internet Measurement Conference
instant metrics on when they lose viewers' interest. This       (Chicago, Illinois, USA, November 04 - 06, 2009). IMC '09. ACM, New
can be expanded into measuring positive attention, as well,     York, NY, 430-441.
                                                                [10] T. Silverston and O. Fourmaux. Measuring P2P IPTV Systems. In
when cross-indexed with positive recommendations in             Proceedings of NOSSDAV, June 2007.
social streams: a user is more likely to recommend an           [11] K. Sripanidkulchai, B. Maggs, and H. Zhang. An analysis of live
item after a particularly interesting scene or "moment" in      streaming workloads on the Internet. In Proceedings of the ACM
the video. When video-on-demand services allow                  SIGCOMM (ICM), Sicily, Italy, Oct 2004.
You can also read