Watch me playing, I am a professional A first study on video game live streaming

M. Kaytoue1, A. Silva1, L. Cerf1, W. Meira Jr.1, C. Ra¨ıssi2

1 2 Belo Horizonte – Brazil Nancy – France

Mining Social Network Dynamics @ WWW 2012 Lyon (France) - 16 April, 2012. Electronic Sports

Watching E-Sport on internet: a new entertainment? Just like traditional sport but with video games Professional commentators, sponsors, tournaments, etc. Professional gamers streaming their games over internet Spectators prefer to watch rather than playing themselves

A new Web community is growing Widely using Web media such as FaceBook, Twitter, etc. and... Live video game streaming platform gaining in popularity Very active, important frequency of events

Watch me playing, I am a professional 2 / 38 N Events and tournaments

Watch me playing, I am a professional 3 / 38 N Social TV

Watch me playing, I am a professional 4 / 38 N Contribution Starting from Twitch.tv audience data From September 29th, 2011 to January 09th, 2012 Every five minutes, get tuples of active streams (date, login, game, description, count, ...)

We propose a first characterization of this community Quantitatively: audience, content length, etc. Qualitatively: What games? Where? etc. Early prediction of the audience Ranking most popular professional gamers

Findings Important for E-Sport actors – With nice perspectives of research

Watch me playing, I am a professional 5 / 38 N Outline

1 A first characterization of the E-Sport community

2 Predicting stream popularity

3 Ranking streamers

4 Conclusion and perspectives

Watch me playing, I am a professional 6 / 38 N A first characterization of the E-Sport community Twitch data acquisition and description

Data From September 29th, 2011 to January 09th, 2012 Every five minutes, get all of active streams and their audience More than 24 millions of tuples Cleaning: missing values, removing illegal streams (1.54%), etc.

field description date The date of crawling of the tuple login Unique identifier of a user/streamer game The game or topic of the stream description A text description of the stream count The number of viewers/spectators watching the stream at a given time

Watch me playing, I am a professional 7 / 38 N A first characterization of the E-Sport community Dataset Summary

Period of analysis Sept 29, 11 - Jan 9, 12 #timestamps 28,292 (832 missing) #logins 129,332 #games 17,749 #tuples 24,018,644 #illegal tuples 369,470 (1.54%) #sessions 1,175,589 #views 27,120,337 Length streamed 215.3 years Length watched 9,622.4 years

Watch me playing, I am a professional 8 / 38 N A first characterization of the E-Sport community Views along the weeks (When?)

70000 1600 viewers streamers

60000 1400

50000 1200

40000 1000 avg nb of viewers

30000 800 avg nb of streamers

20000 600

10000 400 Sun Mon Tue Wed Thu Fri Sat Sun

Watch me playing, I am a professional 9 / 38 N A first characterization of the E-Sport community Geographic distribution (Where?)

Watch me playing, I am a professional 10 / 38 N A first characterization of the E-Sport community Top 20 most popular games (What?)

Game Audience Release StarCraft II 35.05% July 2010 Heroes of Newerth 8.89% May 2010 League of Legends 8.19% Oct. 2009 World of Warcraft 6.24% Nov. 2004 : BO 3.88% Nov. 2010 Street fighter 4 3.26% Apr. 2010 Star Wars (TOR) 2.98% Dec 2011 The Elder Scrolls 2.36% Nov. 2011 MineCraft 2.03% Nov. 2011 Rage 1.98% Oct. 2011 Marvel vs. Capcom 3 1.67% Feb. 2011 (beta) 1.55% Sep. 2011 Battlefield 3 1.39% Oct. 2011 Warcraft III 1.22% July 2002 Halo: Reach 1.20% Sept. 2010 Mario Kart 7 1.18% Dec. 2011 Dark Souls 1.10% Oct. 2011 Zelda SS 1.05% Nov 2011 Gears of War 3 0.93% Sept. 2011 Counter-Strike S 0.89 % Nov. 2004 Others 12.95%

Watch me playing, I am a professional 11 / 38 N A first characterization of the E-Sport community Local game popularity (What?)

Heroes of Newerth Zelda World of Warcraft Warcraft III The Elder Scrolls Mario’s Street Fighter Star Wars Starcraft II Rage MineCraft Marvel vs. Capcom League of Legends Halo Counter-Strike Gears of War Dota Dark Souls

% of daily audience Call of Duty Battlefield

Time (days)

Watch me playing, I am a professional 12 / 38 N A first characterization of the E-Sport community Major E-Sport events (What?)

70000

60000

50000

40000

30000

20000 Oct. 11 Nov. 11 Dec. 11 Jan. 12

IEM N-Y Home Story Cup MLG Orlando NASL S2 Finals IGN Pro League NE League S2 Grand Finals DreamHack Winter 12 hours for charity Blizzard Cup #views

Watch me playing, I am a professional 13 / 38 N A first characterization of the E-Sport community Stream and Streamer characteristics

1 1 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.3 0.4 cumulative (%) 0.2 cumulative (%) 0.3 0.1 0.2 0 0.1 100 101 102 103 104 105 100 101 102 103 104 105 106 duration (min) duration (min) (a) Stream (b) Streamer Duration of streams and aggregate duration of streamers 1 1

0.8 0.8

0.6 0.6

0.4 0.4

agregate views 0.2 agregate views 0.2

0 0 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 stream rank streamer rank (c) Stream (d) Streamer Stream and streamer audience

Watch me playing, I am a professional 14 / 38 N 1 A first characterization of the E-Sport community

2 Predicting stream popularity

3 Ranking streamers

4 Conclusion and perspectives Predicting stream popularity Motivation Current Twitch recommendation strategy New and interesting streams may take too long (or even never) to become visible

Watch me playing, I am a professional 16 / 38 N Predicting stream popularity Motivation

Streaming sessions have a highly skewed popularity distribution, short duration, and slow popularity evolution.

0.3 1 1 average session for the top-100 streamers 0.9 0.25 0.8 0.8 0.7 0.2 0.6 0.6

0.5 0.15 0.4 0.4

0.3 0.1 cumulative (%) agregate views 0.2 0.2 0.1 0.05

0 0 proportion of the overall maximal popularity 0 1 2 3 4 5 0 10 20 30 40 50 60 70 80 90 100 10 10 10 10 10 10 0 0 2 4 6 8 10 12 14 16 stream rank duration (min) hours since the beginning of a session (e) (f) (g)

Stream popularity, duration and popularity evolution

Watch me playing, I am a professional 17 / 38 N Predicting stream popularity Idea

Predicting popularity using initial popularity records

105 105

104 104

103 103

102 102

101 101 popularity after 1 hour popularity after 1 hour 100 100 100 101 102 103 104 105 100 101 102 103 104 105 popularity after ti minutes popularity after ti minutes

(h) ti = 5 min. (i) ti = 30 min.

Correlation between stream popularity after ti minutes and 1 hour

Watch me playing, I am a professional 18 / 38 N Predicting stream popularity

Correlation Varying ti

Correlation between popularity after ti minutes and 1 hour

1 14

0.9 13 12 0.8 11 0.7 correlation 10 0.6

corr. 9 mean squared error ε 0.5 8 5 10 15 20 25 30

ti (min)

Watch me playing, I am a professional 19 / 38 N Predicting stream popularity Prediction Model

Model

log(pop(tf )) = β0 + β1 log(pop(ti )) + 

Predicted vs. actual (based on popularity after ti minutes)

105 105

104 104

103 103

102 102

101 101

0 0 actual popularity after 1 hour 10 actual popularity after 1 hour 10 100 101 102 103 104 105 100 101 102 103 104 105 predicted popularity after 1 hour predicted popularity after 1 hour

(j) ti = 5 min. (k) ti = 30 min.

Watch me playing, I am a professional 20 / 38 N Predicting stream popularity

MSE Varying ti

MSE for different values of ti (minutes)

1 14

0.9 13 12 0.8 11 0.7 correlation 10 0.6

corr. 9 mean squared error ε 0.5 8 5 10 15 20 25 30

ti (min)

Watch me playing, I am a professional 21 / 38 N 1 A first characterization of the E-Sport community

2 Predicting stream popularity

3 Ranking streamers

4 Conclusion and perspectives Ranking streamers Why rank streamers?

Interesting for Spectators: Who to watch? Sponsors: Who to support? Teams: Who to recruit? Gamers: Is my rival doing better? Game editors: Is my game more popular than my concurrents?

Watch me playing, I am a professional 23 / 38 N Ranking streamers Comparing two streamers

Audience depends of other streams active at the same time Comparison of two streamers when they broadcast together

Example On Nov. 10 19:00, WhiteRa is preferred to EG.IdrA. They are not comparable with Mill.Stephano.

crawl time Oct. 29 16:30 Oct. 29 16:35 Nov. 10 19:00 EG.IdrA 1950 6350 1020 Mill.Stephano 4450 3680- WhiteRa 935 2301 4535

Watch me playing, I am a professional 24 / 38 N Ranking streamers Challenge

Difficulty Raw audience is not a good measure of popularity because of: daily/weekly variations of the number of viewers and sessions; variations of the number of viewers along a session.

Idea for aggregating the preferences Consider the streamers as candidates, the crawl points as voters and apply a Condorcet method that is known to be good for ranking: Maximum Majority Voting.

Watch me playing, I am a professional 25 / 38 N Ranking streamers Ranking the pairs of streamers

Three criteria with the following precedence:

c1 How often the first streamer is preferred to the second;

c2 How often they have the exact same popularity;

c3 How often they broadcast at the same time.

c1 c2 c3 (EG.IdrA,WhiteRa) 0.9615 0 156 (EG.IdrA,Mill.Stephano) 0.9 0 20 (WhiteRa,Mill.Stephano) 0.7829 0 175

Watch me playing, I am a professional 26 / 38 N Ranking streamers Building an acyclic directed graph

Until all ranked pairs are processed:

1 Add all tied pairs as edges;

2 For every newly added edge, decide the existence of a cycle involving it;

3 Remove those involved in a cycle;

4 Go to 1.

Watch me playing, I am a professional 27 / 38 N Ranking streamers Resulting graph

Watch me playing, I am a professional 28 / 38 N Ranking streamers Results with Top-100 streamers

Focusing on eight StarCraft II players

Web poll (# votes) Simple ranking (pos.) Condorcet WhiteRa (11,112) EG.IdrA (20) EG.IdrA Mill.Stephano (9,192) WhiteRa (21) Mill.Stephano EG.IdrA (6,746) Liquid’Ret (31) EG.HuK EG.HuK (5,050) EG.HuK (32) WhiteRa Liquid‘HerO (2,160) Mill.Stephano (33) Liquid‘HerO Liquid’Sheth (846) Liquid‘HerO (53) QxG.SaSe QxG.SaSe (833) Liquid’Sheth (72) Liquid’Sheth Liquid’Ret (684) QxG.SaSe (91) Liquid’Ret

Watch me playing, I am a professional 29 / 38 N 1 A first characterization of the E-Sport community

2 Predicting stream popularity

3 Ranking streamers

4 Conclusion and perspectives Conclusion and perspectives Conclusion Characterization of a new Web community Gathered around social TV (Twitch.tv) Quantitative and qualitative characterization Popular tournaments and releases translate into audience Early prediction of future audience of a stream Ranking popular players via a Condorcet method

A particular interest For the actors of this community (spectators, pro-gamers, sponsors, game publishers, etc.) For the research community (social network, data-mining, social sciences, etc.)

Watch me playing, I am a professional 31 / 38 N Conclusion and perspectives Going further into the characterization

A community per se accommodated with Web technologies, intensively using Web media like Facebook, Twitter, YouTube, and very active, making it an interesting study case for researchers.

Further work A better characterization, including other media/data Formally define entities, relations, dimensions, etc

Watch me playing, I am a professional 32 / 38 N Conclusion and perspectives Examples

Propagation Data: Facebook and Twitter streaming announcements Question: How does it propagate into audience?

Network dynamics & Popularity Data: List of IRC users logged in and watching a stream Question: are spectators structured into (evolving) sub-communities? Question: Can we translate spectator moving from a stream to another into popularity?

Watch me playing, I am a professional 33 / 38 N Conclusion and perspectives Examples Popularity: a point of view, depends on several factors Data: Twitch audience, chat session (sentiment analysis) Data: Forum fan-club, e.g. TeamLiquid.net Data: Official season ranking Data: Records of ladder games, e.g. A won against B on day C Question: How/can “Skylines” determine best players? Question: Can we early predict rising/dying stars?

Personal recommendation Data: Twitch data Question: How to recommend an interesting and unknown stream for a spectator?

Watch me playing, I am a professional 34 / 38 N Conclusion and perspectives Examples Facebook, tweets, IRC events, etc. NLP, sentiment analysis (each game has a specific vocabulary) Graph-mining, network analysis

Watch me playing, I am a professional 35 / 38 N Conclusion and perspectives Examples

Artificial Intelligence Abstracting (very!) noisy series of events, without knowing the game state that remains to be approximated

Watch me playing, I am a professional 36 / 38 N Conclusion and perspectives Examples

Watch me playing, I am a professional 37 / 38 N Conclusion and perspectives Thank you! All datasets used for this article are available http://homepages.dcc.ufmg.br/~kaytoue/

Other datasets [email protected]

Watch me playing, I am a professional 38 / 38 N