Content uploaded by Gareth Tyson
Author content
All content in this area was uploaded by Gareth Tyson on Dec 13, 2015
Content may be subject to copyright.
Behind the Game: Exploring the Twitch Streaming
Platform
Jie Deng, Felix Cuadrado, Gareth Tyson, Steve Uhlig
Queen Mary University of London, UK
email: {j.deng,felix.cuadrado,gareth.tyson,steve.uhlig}@qmul.ac.uk
Abstract—Twitch is a streaming platform that lets users
broadcast their screen whilst playing games. People can share
their game experience and interact with others in real time.
Twitch has now become the fourth largest source of peak Internet
traffic in the US. This paper explores the unique nature of this
platform over a 11 month dataset. We find that Twitch is very
different to existing video platforms, with a small number of
games consistently achieving phenomenal dominance. We find a
complex game ecosystem combining consistently popular games
over years, newly released games enjoying bursts of popularity,
and even old games appearing on the platform. Despite a strong
skew of views across channels, the top ranked channels, although
taking a significant share of the viewers, exhibit unexpectedly
high churn. The reason behind this churn lies within another
unique feature of this ecosystem, namely tournaments, live events
that last for a limited amount of time but are capable of attracting
a huge share of views when they take place, as well as dominate
the views of the related games. Overall, our work reveals a
complex and rich ecosystem, very different from existing user
generated content platforms.
I. INTRODUCTION
Video games are a form of entertainment enjoyed by
a diverse, worldwide consumer base. Traditionally, gaming
has been a pastime enjoyed by those who choose to play.
Recently, however, gaming has become a spectator form of
entertainment. Major eSports tournaments such as DreamHack
are enjoyed by millions of online viewers, yet have seen
little attention from mainstream media outlets: enter Twitch.
Twitch is a large-scale video streaming platform used (almost)
exclusively for live game broadcasting. It allows broadcasters
to construct channels through which they can stream their
gameplay to the world. Popular types of channels include am-
ateurs broadcasting their gameplay, competitive eSports tour-
naments with commentaries, coaching game sessions, charity
marathon events, and, finally, experimental large-scale coop-
erative events where games are played collectively. Twitch’s
popularity is undeniable, recently being reported as the 4th
largest website in the U.S. by peak traffic [3].
Despite this, we are yet to gain a comprehensive under-
standing of how users, games and broadcasters operate in this
new environment. Twitch holds a number of novel features. Of
most interest is the introduction of a new object of interest,
the game. Whereas in traditional content services, channels
tend to stream unique content, in Twitch the same games
are streamed by many different people. This unusual property
raises many questions in terms of how viewers are spread
amongst these different games. In this paper we expose the
Twitch platform by analyzing live viewing figures collected
over an 11 month period. First, we detail our dataset and
provide insight into the growing scale of Twitch (§II). We
witness significant viewing peaks, approaching one million
simultaneous viewers. We show the rich ecosystem of games
streamed through Twitch (§III), exposing the impact of game
features such as the release date and the genre. We discover a
significant number of gaming events being broadcast through
the platform (§IV) bringing Twitch closer to a live sports
platform than a user-generated content one. Finally we look
at the individual broadcasters (§V), revealing an extremely
skewed popularity distribution, and significant dynamics in
popularity.
II. OVE RVIEW OF TWITCH AND DATAS ET
Twitch is a live streaming video platform focusing on
gaming. It allows users to broadcast themselves playing games
to others, who tune in via a web interface. Streams include
playthroughs of games by amateur users, and large-scale
broadcasts of eSports competitions. Users in Twitch can take
one of two roles, broadcaster or viewer. A broadcaster is
somebody who streams their game play via a dedicated
channel, whilst a viewer is somebody who watches the chan-
nel. Each streamer is limited to one live channel, which is
only online for a fixed period of time when the player is
broadcasting. To facilitate communication, the channels are
enabled with an in-built chat room to allow the users (both
broadcasters and viewers) to interact with each other.
To study Twitch, we have collected a dataset spanning
11 months from February 2014 to December 2014. This
was collected using the public Twitch REST API, which we
contacted repeatedly to extract available metadata. For each
channel, we retrieved the game being played, the number of
viewers, the title of the channel and the broadcaster’s name.
We repeated this every 15 minutes to build a time series
of metadata for every channel in the system, totalling 323
million channel samples, which includes 5.2 million unique
broadcasters.
We begin our analysis by inspecting the overall trends
of Twitch as a service. On average across all 15 minute
snapshots, there are 9100 broadcasters feeding 362k viewers.
We have observed significant growth over the measurement
period, with viewers increasing by 25% and channels by 45%
when comparing monthly views. The service shows significant
dynamic behaviour, with August 2014 reaching a peak of 934k
simultaneous viewers.
III. EXPLORING GAMES
The previous section has presented the Twitch platform and
highlighted its overall growth. As a unique aspect of Twitch,
we first investigate games.
A. Which games are played?
A basic and fundamental question to understand Twitch
is what games are actually broadcast? To explore this, we
use the ‘game’ field returned from the Twitch API. We
observe a total number of 127,497 different name strings.
The reason for this high number is that the field can be
set manually by broadcasters, making it unreliable. To obtain
something close to a ground truth, we collected game data
from ’TheGamesDB.net’ [2], an open, online database for
video games, and ’GiantBomb’ [1], an American video game
and wiki. We match exact names (in lower case, removing
spaces) with these two databases to filter out bogus ’game’
field entries. This leaves a surprisingly large number (21k)
of unique games played on Twitch, accounting for 95%
of all views. We decided against more sophisticated name
matching techniques (such as Levenshtein distance) because
of the predominance of game sequels that would be wrongly
resolved.
We next inspect the respective viewing and broadcast figures
attained by these 21k games. We therefore begin by inspecting
which games are most frequently broadcast and viewed in
Twitch. Much like music artists vie for airtime on radio, game
developers may wish to see their games effectively promoted
on Twitch. We observed 41 games that were streamed in
Twitch prior to their release date (often through official
channels from the developers). This suggests that developers
are using Twitch to gather feedback and generate interest on
upcoming games.
Table I presents the top 10 (0.04%) games found in Twitch.
It can be seen that there is a strongly uneven distribution
of viewers and broadcasters across these games. The top
10% of games collect 95% of all viewers (far higher than
seen with other types of content objects, e.g., Youtube [7] or
VoD [4]). Alone, the top 10 covers 64% of all viewers. This
leaves a large number of games to languish with very little
attention: 92.8% (≈19.6k) of the games have fewer than 100
viewers on average. Twitch has therefore become an ecosystem
largely driven by several extremely popular games. Table I
also presents the percentage of time each game spent in the
top 10. We find that these ranks are remarkably stable; the
most notable example is League of Legends, which held the
top rank for 90.4% of our dataset (staying in the top 10 for
99.5%).
Rank Name % of
Viewers
% of
Channels
% of Uniq
Streamers
% in
Top 10
1 League of Legends 29.1 14.9 11.9 99.5
2 DOTA 2 11 3.2 2.8 99.1
3Hearthstone:
Heroes of Warcraft 8 2 3.3 98.9
4Counter Strike:
Global Offensive 6.2 3.9 5.2 91.67
5 Minecraft 3.5 4.2 8.5 86.2
6Starcraft II:
HeartoftheSwarm 3.0 1.0 0.6 68.3
7WorldofWarcraft:
Mists of Pandaria 2.2 2.8 1.9 71.3
8DiabloIII:
ReaperofSouls 1.9 1.9 1.7 33.5
9 DayZ 1.5 1.5 1.6 28.7
10 Call of Duty:
Ghosts 1.0 2.2 4.1 20.7
TABLE I: Top 10 games in Twitch.
B. Game features
The previous section has highlighted that a few prominent
games dominate Twitch. A key question is what are their
characteristics? To explore this, we augment our games with
metadata obtained by the API from GamesDB [2] and Giant-
Bomb [1]. We focus on release date and genre.
1) Release date: A common property held across many
content repositories is user preference for recent releases, as
consumers tend to constantly seek out new stimulation [16],
[19]. We investigate if Twitch shares this property. Specifically,
we inspect if new game releases gain popularity and manage
to steal viewers away from older and more established games.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
Dec
Fraction
Day
2009 ~ 3 month
3 month ~
~ 2009
Fig. 1: Fraction of views across games with different release
times: between 2009 to three months before the observation
point, within three months and games released before 2009.
Figure 1 breaks down the viewings across games split into
3 bins of release date. For the games released every year,
we calculate the number of viewers each game collects in
each snapshot. We set our main boundary in 2009, as all the
top games in Table I were released between that year and
today. The popularity of these games is shown clearly in the
data; games released between 2009 and 3 months before the
data point consistently attain ≈75% of the viewer share. Thus,
game popularity is not as ephemeral as seen in other domains,
which frequently see newly released content at the top of the
rankings [16], [19]. In Twitch, fresh games (i.e., <3 months)
manage to attract only 18% of viewers; this can be contrasted
with YouTube, in which new videos constitute 80% of the top
list [7]. Surprisingly, we also witness a notable share (7%)
for very old games released prior to 2009. This fraction of
the share is bursty, with peaks of 40% of the total platform
views, and significant variation in the individual games that
contribute to these shares. We will revisit this burstiness of
recent games later in the paper.
2) Genres: We further explore the games in Twitch by
classifying them into different genres (based on information
from the GiantBomb API). Figure 2 shows the fraction of
views taken by the top genres across the measurement period.
We find that MOBA (i.e., multiplayer online battle arena) is
the dominant genre, gathering 40% of the views (as expected
with the top 2 games belonging to that genre. Nonetheless,
high variance can be seen with significant changes across the
measurement period. One of the most abrupt changes is the
big drop in the aggregate share of these popular genres at the
beginning of June. This corresponds to the broadcasting of
the “E3 press conferences” (the main video games festival)
on June 9th 2014, which shifted viewers away from the
usual content genres enjoyed by users. It can also be seen
that notable shifts occur with views transferring between
these popular genres; for example, role playing games lost
a significant share of their viewing figures to other genres
in mid-April (this was caused by the fading popularity of a
viral phenomenon, Twitch Plays Pokemon [11]). This is just
one example of the complex interactions between simultaneous
events occurring in Twitch.
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Mar
Apr
May
Jun
Jul
Aug
Sep
Oct
Nov
Dec
Fraction
Day
Others
real-timestrategy
action-adventure
action
mmorpg
role-playing
cardgame
first-personshooter
moba
Fig. 2: Fraction of daily views for each genre.
IV. EXPLORING TOURNAMENTS
The above exploration has revealed a particularly important
type of channel: tournament broadcasts. The real time nature
of Twitch makes it ideal for broadcasting live events taking
place anywhere in the world. This is the case for the burgeon-
ing eSports competitive scene.
We have manually extracted the key events streamed via
Twitch throughout 2014. In total, we have identified 56
events, some lasting multiple days. 53 are eSports tourna-
ments (we left out of this analysis competitive leagues and
preliminary phases); the remaining 3 are a charity event, the
famous E3 press conference, and the TwitchPlaysPokemon
phenomenon [11]. The total number of days from our dataset
with an event running are 150 (47% of the considered days).
Figure 3 presents each event’s share of the overall viewing
figures over time (each bar is an event colour coded by the
game played1). The huge impact of event broadcasters is
undeniable; many exceed 20% of the daily viewers, making
them the top ranked channel (Figure 7). Whereas the average
top ranked channel’s viewing share is 8.9% (daily), this can
increase up to 30.5% during events. In absolute terms we
observe close to a million simultaneous viewers in the whole
platform. Thus, Twitch resembles other live sports platforms,
with spikes during key events [20].
0.0
0.1
0.2
0.3
Feb Apr Jun Aug Oct Dec
Days
Daily share of views
cs:go dota2 e3 hs lol pokemon sc2 sf4
Fig. 3: Fraction of viewing figures collected by tournaments
per day. Tournaments are grouped per game: Counter Strike:
Global Offensive, DOTA 2, E3 (gaming trade show), Hearth-
stone, League of Legends, Pokemon, Starcraft 2, and Street
Fighter 4.
We also inspect the relationship between events and games
(see color code in Figure 3). Tournaments playing League of
Legends (LoL) achieve the highest and most frequent peaks.
Numerous events playing other less popular games also reach
comparable levels of popularity over time. Again, we see that
event broadcasters gather the attention of Twitch viewers, even
if they are streaming not so popular games.
Figure 4 shows a CDF of the share of viewers garnered
by the tournaments (on a per game basis); we take samples
from every 15 minutes. Overall, we observe that tournaments
have a substantial impact on every game (with a lesser extent
for DOTA2, which incorporates and incentivises watching
1We treat E3 as a separate game, although this is actually a trade show.
tournaments through their own platform). In the most extreme
case, for Street Fighter 4, the tournament can reach 100% of
all viewers of that game. The sometimes large share of viewers
captured by tournaments for a given game explains why the
companies behind these games provide so much support for
the tournaments, e.g., in the form of prize money.
Overall, tournaments nicely illustrate the complexity of the
Twitch platform, that lies at the crossing between viewers,
broadcasters, as well as different types of companies (e.g.,
game development and advertising).
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 0.2 0.4 0.6 0.8 1
CDF
Event share of game views per timestamp
dota2
lol
cs:go
hs
sc2
sf4
Fig. 4: CDF of the share of game views gathered by events in
each snapshot.
V. EXPLORING CHANNELS
In this section, we explore the relationship between viewers,
games and channels.
A. Channels 6=Games
We have previously shown that some games (e.g., League of
Legends) are extremely popular in Twitch. Hence, it is natural
to expect a similar skew in channels. Figure 5 presents the
distribution of viewers across channels (x-axis provides the
rank of the channel), confirming that much like with games,
channel popularity is highly skewed towards a few prominent
broadcasters. The top 10 broadcasters alone collect 16% of all
views. In each 15 minute snapshot, the top 1% of channels
collect 70% of the viewers, whilst the top 10% collect 93% of
all viewers. In contrast, the remaining channels get very poor
viewing figures (62% of them below 1 viewer on average).
The majority of broadcasters therefore make little impact on
the overall system — Twitch is a system dominated by a tiny
minority. Figure 5 also provides the distribution of viewers
across subscription channels (whitelisted by Twitch under
the “partners” scheme). Subscription channels provide the
opportunity for viewers to support their favourite broadcasters
by paying an optional monthly fee. Subscription is shared
between Twitch and the broadcaster. We observe from Figure 5
that subscription channels are a representative subset of the
most popular channels in terms of viewing figures and skew.
Consistent with their purpose, we also observe that no low
ranked channel is among the subscription channels.
1
10
100
1000
10000
100000
1e+06
1e+07
1e+08
1e+09
1
10
100
1000
10000
100000
1e+06
1e+07
No. of viewers
Rank (by total number of viewers)
All channels
Subscription channels
Fig. 5: Distribution of viewers per channel.
Now we come to the question of the “relationship” between
games and channels. To answer this, Figure 6 shows how many
games have been played by each channel over its lifespan.
Substantial diversity can be observed, with some broadcasters
playing in excess of 100 games. As such, channels 6=games:
often channels broadcast many different games. Figure 6 also
breaks down broadcasters into different popularity groups; it
can be seen that broadcasters with high viewing figures tend
to play many games. For channels that garner between 100
and 10k viewers, only 30.4% play a single game. Moreover,
roughly half of the popular subscription channels play at least
ten games. That said, curiously, extremely popular channels
(>10k viewers) tend to play fewer games. This indicates that
there is no direct relationship between gaining viewers and the
number of games played. Instead, popularity appears inherent
to the broadcaster, rather than the games they play. This is a
key observation when trying to understand Twitch, particularly
when considering companies wishing to advertise games (e.g.,
by getting popular broadcasters to promote it by playing).
B. Churn in top channels
So far, we have shown that the top 10 channels garner a
significant share of the viewers. We next see how this share
varies over time. Figure 7 shows, for each week, the share of
views accumulated by the top 10 channels. Despite seeing
a significant concentration of views, the total share rarely
surpasses 30% of the weekly views. While Figure 7 might
give the impression that the share of viewers captured by the
channels showed in this top 10 is stable, this is actually not
the case. Indeed, there is significant churn in which actual
channels are those that are ranked within this top 10 over time.
Figure 8 shows a visual representation of the top 10 channels
every week. Each coloured symbol depicts a different channel,
with lines showing how they move up and down the rankings.
The top ranked channels (i.e., ranked first) are extremely
consistent; for example, although it only streams for 80% of
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1 10 100 1000
CDF
Number of games
Viewer < 100
100 < viewer < 10000
viewer > 10000
Subscription channels
Fig. 6: Number of games played by each channel. Channels
are separated into popularity groups by average number of
viewers.
the observed period, ‘riotgames’ is the top channel 60% of the
time. This consistency is unusual considering the short-lived
nature of popularity in other domains, e.g., YouTube. However,
as can be seen thanks to the solid lines that connect every
coloured symbol, most of the channels exhibit a significant
amount of churn both in terms of their presence within the
top 10, as well as their actual rank. The “randomised” look
of Figure 8 visualises well the amount of churn taking place
within top ranked channels.
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan
Fraction
Week
Fig. 7: Platform share of the top 10 channels each week by
number of viewers.
A closer inspection of the weekly top channel over the
measurement period reveals that 95% of the time the top
ranked channel is riotgames, the official channel for League
of Legends, dedicated to broadcasting official eSports tour-
naments. Individual broadcasters (as opposed to corporations)
regularly appear in the top list of channels, although clearly
the most popular channels are broadcasted live events. Indeed,
Twitch is a natural platform for this type of content (as viewers
can interact while they watch, and live streaming allows to
broadcast the event as it happens).
1
2
3
4
5
6
7
8
9
10 Apr Jul Oct Jan
week
rank
Fig. 8: Top 10 channels each week by number of viewers.
Lines connect reoccurring instances of the same channel.
VI. IMPLICATIONS
We have observed several unique characteristics that dif-
ferentiate Twitch from other large-scale streaming services.
Twitch constitutes a novel form of multimedia, introducing
the concept of game as a new multimedia object, possess-
ing unique popularity characteristics (distinct from traditional
channels/broadcasters). Exploiting the knowledge introduced
by these games is a ripe area of further exploration. For
instance, use of game information for things like recommender
engines would be highly beneficial. Similarly, the observation
that many users watch the same games (potentially played
by different streamers) may indicate that such recommender
engines could easily guide users to nearby (low-cost) stream-
ers, rather than distant (high-cost) streamers playing the same
game [14]. A key finding of our work is the preference users
have for particular users, which suggests that this would have
to be done with caution: players are certainly not all equal.
Unlike many other streaming services, Twitch also benefits
from highly predictable popularity trends regarding its flash
crowds. This is because Twitch flashcrowds are almost always
driven by scheduled tournaments. In contrast, open video
services (e.g., YouTube) are often left unaware of new up and
coming events that may generate flashcrowds. Naturally, this
is driven by the existence of many different unknown video
providers and genres. This could make Twitch’s infrastructural
provisioning much more straightforward than other platform’s.
Similarly, outside of tournaments, a relatively small number
of extremely popular broadcasters share >90% of online
viewers. This means that predicting the behaviour of this
small number of broadcasters could lead to similar benefits
when provisioning infrastructure. Such broadcasters could
even be asked to release schedules, so Twitch could know
in advance how to provision its infrastructure. The value of
this predictability cannot be underestimated.
Twitch delay settings allow streamers to configure the delay
of a stream before transmission. This feature was introduced
to prevent multiplayer cheating, but could also be exploited
by the infrastructure. Specifically, the delay parameter could
be used to enable caching and staggering of delivery across
users. This would be, in essence, consensual buffering.
Another key implication is the impact that Twitch has
on the games industry more generally. Curiously, we find
little correlation between traditional games ratings and the
popularity of games on Twitch. This suggests that users
watching and playing games may have different needs. This
has clear ramifications for the games industry. Most notably, it
introduces a powerful means to gather rapid feedback. This is
clearly something that is already being explored, as we found
41 games pre-released on Twitch, as part of the promotion
activity. We envisage games will increasingly be designed with
Twitch-like broadcast in-mind. We already see this with many
prominent games having in-built support for Twitch. Finally,
the growing popularity of Twitch should be treated as a wake-
up call for TV broadcast outlets. Broadcast is an extremely
efficient medium for popular content and, as such, we argue
it is only a matter of time before gameplay content becomes
commonplace on TV broadcast, much like other sports.
VII. REL ATED W OR K
Video gaming has a history spanning decades. Recently,
researchers have turned their attentions to online gaming, look-
ing at its evolution [10]. Furthermore, online social gaming has
recently emerged as a hot topic, as it integrates the fields of
gaming with that of social networking [9], [6], [13]. This work
is very different to our own, though, as we focus on the nature
of streaming games, rather than playing them.
It could be argued that Twitch is more closely related to
general video streaming platforms, particularly user generated
content (UGC) repositories. Many studies have been devoted
to other UGC platforms [5], [7], [19]. Our work is orthogonal
to these, as we show the live broadcasting nature of Twitch
makes it fundamentally different to these platforms.
A small set of researchers have started look into Twitch
from different perspectives. Most notably, [12] explored the
early stages of Twitch (in 2012), finding channel viewing
figures were highly predictable by looking at its early viewing
figures. We have explored the actual content (i.e., games and
tournaments) being broadcast by channels. This has allowed us
to shed light on the underlying driving factors of popularity on
Twitch. Other researchers have recently explored the delivery
infrastructure of Twitch; for instance, measuring the traffic
generated by certain channels [8], [17], [18], mapping the
video delivery infrastructure [21] and building models of user
chat interactions [15].
VIII. CONCLUSION
In this paper we have explored the most popular game
streaming platform in the world, Twitch. We have uncovered in
the paper that the popularity of game streams possesses unique
characteristics distinct from traditional channels, broadcasters
and games. We have uncovered a complex ecosystem, with
multiple types of content competing for the attention of users.
Newly released games gather a relatively small part of the
views, particularly when compared to the ephemeral nature of
content in other UGC platforms.
A significant part of Twitch activity is centered on live
gaming events, some of which dominate the views when they
are taking place. From this angle, Twitch resembles traditional
TV sports broadcasting. These events generate predictable
flash crowds (as they follow a schedule), gathering millions
of concurrent viewers.
REFERENCES
[1] Giantbomb.com. www.giantbomb.com/api.
[2] Thegamesdb.net. thegamesdb.net.
[3] Twitch ranked 4th in peak internet traffic. http://blog.twitch.tv/2014/02/
twitch-community- 4th-in- peak-us- internet-traffic/.
[4] H. Abrahamsson and M. Nordmark. Program popularity and viewer
behaviour in a large tv-on-demand system. pages 199–210, New York,
NY, USA, 2012. ACM.
[5] P. Ameigeiras. Analysis and modelling of YouTube traffic. Transactions
on Emerging Telecommunications Technologies, (2):360–377, 2012.
[6] J. Blackburn and H. Kwak. STFU NOOB!: predicting crowdsourced
decisions on toxic behavior in online games. Proc. of WWW, pages
877–887, 2014.
[7] M. Cha, H. Kwak, P. Rodriguez, Y.-y. Ahn, and S. Moon. I tube,
you tube, everybody tubes: analyzing the world’s largest user generated
content video system. Proc. of ACM IMC, 2007.
[8] F. Chen, C. Zhang, F. Wang, and J. Liu. Crowdsourced live streaming
over the cloud. arXiv preprint arXiv:1502.06314, 2015.
[9] T. Chung, J. Han, D. Choi, and T. Kwon. Unveiling group characteristics
in online social games: a socio-economic analysis. Proc. of WWW, pages
889–900, 2014.
[10] S. Gallagher and S. Park. Innovation and competition in standard-based
industries: a historical analysis of the US home video game market.
IEEE Engineering Management, 49(1):67–82, 2002.
[11] A. Hern. Twitch plays pok´
emon: live gamings latest big hit. The
Guardian, 2014.
[12] M. Kaytoue, A. Silva, L. Cerf, W. Meira Jr, and C. Ra¨
ıssi. Watch me
playing, i am a professional: a first study on video game live streaming.
In Proc. of WWW Companion, pages 1181–1188, 2012.
[13] R. Kowert and J. A. Oldmeadow. 965-976. Computers in Human
Behavior, June 2014.
[14] D. K. Krishnappa, M. Zink, C. Griwodz, and P. Halvorsen. Cache-
centric video recommendation: An approach to improve the efficiency
of youtube caches. ACM Trans. Multimedia Comput. Commun. Appl.,
11(4):48:1–48:20, June 2015.
[15] G. Nascimento, M. Ribeiro, L. Cerf, N. Cesario, M. Kaytoue, C. Raissi,
T. Vasconcelos, and W. Meira. Modeling and analyzing the video game
live-streaming community. In Proc. of IEEE LA-WEB, pages 1–9. IEEE,
2014.
[16] G. Nencioni, N. Sastry, J. Chandaria, and J. Crowcroft. Understanding
and decreasing the network footprint of over-the-top on-demand delivery
of tv content. In Proc. of WWW, pages 965–976, 2013.
[17] K. Pires and G. Simon. Dash in twitch: Adaptive bitrate streaming in
live game streaming platforms. In Proc. of VideoNext, pages 13–18,
2014.
[18] K. Pires and G. Simon. Youtube live and twitch: a tour of user-generated
live streaming systems. In Proc. of ACM MMSys, pages 225–230, 2015.
[19] G. Tyson, Y. Elkhatib, N. Sastry, and S. Uhlig. Demystifying porn 2.0:
a look into a major adult video streaming website. In Proc. of ACM
IMC, pages 417–426, 2013.
[20] H. Yin, X. Liu, F. Qiu, N. Xia, C. Lin, H. Zhang, V. Sekar, and G. Min.
Inside the bird’s nest: measurements of large-scale live vod from the
2008 olympics. In Proc. of ACM IMC, pages 442–455, 2009.
[21] C. Zhang and J. Liu. On crowdsourced interactive live streaming: a
twitch. tv-based measurement study. In Proceedings of the 25th ACM
Workshop on Network and Operating Systems Support for Digital Audio
and Video, pages 55–60. ACM, 2015.
