iscool entertainment @big data paris

21
Big Data + Social + Games @Is Cool 3/24/12 TITRE DOCUMENT

Upload: iscoolent

Post on 05-Jul-2015

432 views

Category:

Business


4 download

DESCRIPTION

"Big Data + social + games @IsCool"A presentation by IsCool Entertainment at the Big Data Congress Paris.

TRANSCRIPT

Page 1: IsCool Entertainment @Big Data Paris

Big  Data    +  Social  +  Games  @Is  Cool    

3/24/12  TITRE  DOCUMENT  

Page 2: IsCool Entertainment @Big Data Paris

IsCool  Entertainment  

!   Social  game  publisher  based  in  Paris,  France  

!   #1  French  publisher  in  terms  of  audience  (450k  Daily  AcKve  Users)  &  revenue  

!   2.8  Million  Fans    !   80  employees  !   €9.1  million  revenue  in  2010  !   4  live  applicaKons  on  Facebook  

Florian  DoueTeau    CTO  @fdoue?eau  

Agenda  •  What  do  we  do  Social  Gaming  •  What  kind  of  (Big)  AnalyKcs  we  do  Lots  •  How  we  do  it    Hadoop,  Python,  R,  Tableau,  Gephi  and  stuff…  

Page 3: IsCool Entertainment @Big Data Paris

Is  Cool  Games  

Is  Cool,  Delirious  CollecKble    

Game  

Absolute  Solitaire,  The  best  solitaire  game    available  online  

Temple  Of  Mahjong,  Collect,  Play,  Exchange  

Belote  MulKjoueur,  Play,  Win,  Meet  

Page 4: IsCool Entertainment @Big Data Paris

Games  &  Virtual  Goods      

!   Play  the  Game  &  Gain  some  virtual  goods  

!   Play  again  &  Gain  more  ! Collaborate  with  other  players  

&  Gain  More  !   ….  ! Possibly  buy  §  To  grow  quicker  §  To  help  others  

Page 5: IsCool Entertainment @Big Data Paris

Virtual  Goods    Virtual  Economy  

!   Virtual  Goods  Must  not  be  too  easy  to  get  §  The  game  would  not  be  fun!    §  No  moneKzaKon  

!   Virtual  Goods  must  not  be  hard  to  get    §  People  would  churn  because  of  

frustraKon!    !   Virtual  Goods  can  usually  be  

traded  between  players  !   Virtual  and  actual  “Price”  of  a  

good  

Let’s  Trade  1  Watch  against  3  Hammers  

Page 6: IsCool Entertainment @Big Data Paris

Why  is  this  Big  Data  ?    

!   Number  of  object  transacKons  per  day  §  NYSE      3,600,000,000    §        1,600,000,000    

§        1,500,000,000    

§  IsCool      1,400,000,000  

§         860,000,000  

§  CAC  40    142,500,000  

9,8  TB  Data  to  analyze        

18  Million  user-­‐generated  acKons  per  day      7  Billion  per  year.      

Page 7: IsCool Entertainment @Big Data Paris

The  Real  Big  Data  Challenge    Collaborate  for  collecQve  insights  

data  scienKst?    

what  metrics?  

Real-­‐Kme?  

Game  Designer  PerspecKve  :    Nice  Charts  ?    

Programmers’  PerspecKve  :    Log  Files  &  Work  ?    

Business  Guy  PerspecKve:    Revenue  Forecast  ?    

BI  Veteran:    Schema  DefiniKon  ?    

Page 8: IsCool Entertainment @Big Data Paris

Specifics  of  Game  AnalyQcs  

!    Virtual  Goods  §  We  are  the  Factory  AND  the  

Shop,  and  most  of  the  products  are  free.    

!   Social  Networks  §  Network  effects  are  key    

!   Games    §  The  product  changes  EVERY  day  !  §  Sudden  wage  of  unexpected  

players  from  Guatemala  !    §  People  try  to  cheat  !      

Page 9: IsCool Entertainment @Big Data Paris

Use  Case  1:  Engagement    Drivers  

!   StaKsKcal  Mesaure  of  Engagement    §  Visit  Frequency,  DAU  /  

MAU  !   Analyze  Engagement  Drivers  §  Use  of  Features  ?    §  Demographics  ?    §  How  does  it  relate  in  Kme  

with  moneKzaKon  ?    §  ….    

3/24/12  

Page 10: IsCool Entertainment @Big Data Paris

Understanding  Engagement  -­‐    Results  

!   Establish  class  of  users  with  different  engagement  profile  and  use  of  features  

Page 11: IsCool Entertainment @Big Data Paris

Understanding  Engagement  –  Benefits  

!   Adapt  the  features  correlated  with  strong  engagement  and  interesKng  profile  so  that  they  can  easily  be  accessed  by  other  players  

3/24/12  

Page 12: IsCool Entertainment @Big Data Paris

Use  Case  2  :  Understanding  Users  as  a  whole  

!   10  Million  Nodes    !   Around  1  Billion  Edges                  

! How  does  the  graph  evolve  in  Kme  ?    

!   What  are  the  communiKes?  !   Leaders  ?    !   CorrelaKon  with  engagement,  

virality  ,  etc..  ?                  

Page 13: IsCool Entertainment @Big Data Paris

Understanding  Users  as  a  Whole  –  Clusters  and  Graphs  

A  very  large  community  

Some  mid-­‐size  communi6es  Lots  of  small  clusters    mostly  2  players)  

!   Specific  communiKes  in  the  graph  

!   CorrelaKon  between  community  size  and  engagement  /  virality    

!   DetecKon  of  paTerns  §  2  players  paTerns  §  Family  play  §  Group  Play  §  Open  Play  (language  

community)    

           

Page 14: IsCool Entertainment @Big Data Paris

Understanding  Users  as  a  Whole  -­‐  Benefits  

3/24/12  

!   Cluster-­‐oriented  Community-­‐Management  §  Engage  with  a  community  as  

a  whole  as  much  as  possible  

!   Nurture  communiKes    §  Make  communiKes  grow  unKl  

they  reach  a  criKcal  mass  §  Reduce  language  barrier  to  

help  community  aggregaKons  !   DetecKon  of  “opinion  leaders”  

   

Page 15: IsCool Entertainment @Big Data Paris

Use  Case  3  :  Long  Terms  effects  of  a  feature  

3/24/12  TITRE  DOCUMENT  

!   Are  players  using  the  new  feature…  §  Happy  with  it  ?    §  More  engaged?  §  Generate  more  virality  ?  §  etc….  

 !    A/B  Tests  §  Some  features  can  be  A/B  tested  §  …and  some  cannot  !    §  How  to  measure  the  uplio  ?    

!   Complexity  §  MulKple  variable  to  observe  (other  

features,  history,  and  20  more  ….)  §  Long  term  non  local  effect  (game  

economics)    

 

Page 16: IsCool Entertainment @Big Data Paris

Long  Terms  Effects  of  a  feature  -­‐  Results  

3/24/12  

!   Adapt  game  rules  to  fit  most  of  the  players  §  No  InflaKon    §  But  maintain  Growth  !!    

Page 17: IsCool Entertainment @Big Data Paris

How  did  we  do  that  ?    

In  the  past  4  years  ….  •  Tools  changed  •  Scale  changed  •  Focus  Changed  

Technological  Offering  •  Commercial  /  Open  Source  ETL    •  Commercial  BI  VisualizaKon  Sooware  

•  Commercial  /  Open  Source  databases  (column  stores)  

• …  

• Big  Data  Approach  

2010-­‐2011    

• BI  Approach  

2009-­‐2010  

• Basic  Approach  

2008-­‐2009  

Page 18: IsCool Entertainment @Big Data Paris

What  we  learned  

Diversity  

RelaKvity        

CollaboraKon  

No  Hadoop+R  Magic  

Windows  /  Linux  ?      

What  is  real  budget  ?    

VisualizaQon    is  more  important  than  precision  

Do  you  have  data  mining  experts  (yes/no)  ?    

Do  you  have  scalability  experts    ?    

Cloud  or  on-­‐premise  ?  

Do  you  want  anybody  to  play  with  the  data  ?    

No  XYZ  Magical  Product      

Page 19: IsCool Entertainment @Big Data Paris

!   Ad-­‐hoc  -­‐  Datamining  tools  §  To  Discover  new  trends    §  Ad-­‐hoc  analyKcs  §  Graph  AnalyKcs  

!   Week-­‐to-­‐Week  -­‐  Datawarehousing  §  Detailed  Business  Metrics  §  Virtual  Economy  Modeling  §  Long-­‐term  behaviours  §  Business  Level  Visibility  

AdapQve  AnalyQcs  

!   Day-­‐To-­‐Day  -­‐  SaaS  AnalyKcs  Plarorms  §  For  common,  business  metrics  

 (virality,  traffic,  engagement)  §  Corporate  Level  Visibility  §  Day-­‐to-­‐day    

Page 20: IsCool Entertainment @Big Data Paris

Internal  Data  Warehousing    

• +Direct  connecKon  to  the  database    

• +Excel  fan  biz  guy  can  use  it  with  no  training  !    

VisualizaKon      (Tableau  Sooware)  

• Free  (as  beer)    • Good  performance  for  analyKcs  tasks  on  a  few  hundreds  million  lines  (  SELECT  …  GROUP  BY  …  ORDER  …  )      

• Featured  and  limited  performance  compared  to  commercial  Column  Stores                                            

Columnar  Database  (Infinidb)  

• Pure  Python  ETL    • Good  integraKon  with  AWS/  S3    

• Easy  to  integrate  in  our  development  environment    

ETL  (PyBabe)  

• Used  to  reduce  the  amount  of  informaKon  :  10  GB  a  day  =>  1GB  a  day  

• High  cost  of  development  for  business-­‐related  processing                      

MapReduce  (Hadoop/Hive)    

Page 21: IsCool Entertainment @Big Data Paris

?