23
LEARNING HOW HUMANS AND NONHUMANS INTERACT WITH DIGITAL ADS H2O World, November 10, 2015 Sergei Izrailev SVP, Data Science [email protected]

H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

Embed Size (px)

Citation preview

Page 1: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

LEARNING  HOW  HUMANS  AND  NON-­‐HUMANS  INTERACT  WITH  DIGITAL  ADS  

H2O  World,  November  10,  2015  

Sergei  Izrailev  SVP,  Data  Science  [email protected]  

Page 2: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

1  

•  About  Integral  Ad  Science  - What  we  do  -  About  our  data  -  Tech  stack  

•  Non-­‐Humans:  Ad  Fraud  - Overview  - DetecKon  and  prevenKon  

•  Human  InteracKon  - Did  anyone  see  the  ad?  - What  metrics  are  important?  - How  can  we  help  the  brand’s  message  to  be  heard?  

H20  World,  11/10/2015  

AGENDA  

Page 3: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

2  

DIGITAL  ADVERTISING  

MARKETING MESSAGE “Right  Offer”  

1 AUDIENCE TARGETING “Right  Person”  2

MEDIA PLACEMENT “Right  Place”  3

CAUSAL IMPACT & USER ENGAGEMENT

“ROI”  

Page 4: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

•  Brand  Safety  -  The  brand  is  protected  

•  Fraud  DetecKon  -  You  only  pay  for  ads  delivered  to  humans  

•  Geo  VerificaKon  -  The  ads  are  delivered  in  the  right  market  /  region  

•  Viewability  -  A  person  actually  has  a  chance  to  see  the  ad  and  has  the  opportunity  to  engage  

WE  HELP  BRANDS  ENSURE  THEIR  STORIES  ARE  HEARD  

   

     3  H20  World,  11/10/2015  

ESTABLISH  A  BASELINE  –  AND  GO  BEYOND  

ELIMINATE  ADS  THAT  HAVE  NO  CHANCE  OF  DELIVERING  A  MESSAGE  

Page 5: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

4  

•  ReporKng  - What  happened    

•  Blocking  the  ad  from  showing  in  bad  places  -  Before  it’s  too  late  

•  TargeKng  good  environments  -  Before  you  buy  

•  CreaKng  and  packaging  good  environments  -  Before  you  sell    

H20  World,  11/10/2015  

HOW  THE  DATA  IS  USED  

Page 6: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

H20  World,  11/10/2015  

   

     5  

DATA  COLLECTION  AND  PROCESSING  STACK  

Page 7: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

NON-­‐HUMANS:  AD  FRAUD  

Page 8: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

H20  World,  11/10/2015  

   

     7  

AD  FRAUD  NEARLY  ALL  AD  FRAUD  IS  CAUSED  BY  BOT  ACTIVITY  

Ad  Stacking  Placing  mulKple  ads  on  top  of  one  other  in  a  single  ad  placement,  with  only  the  top  ad  in  view  

Illegal  Bots  Compromised  computers  with  breached  security  defenses  conceded  to  a  third  party  

Pixel  Stuffing  Stuffing  an  enKre  ad-­‐supported  site  into  a  1x1  pixel  AD  

Page 9: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

H20  World,  11/10/2015  

   

     8  

FRAUD  DETECTION  

facebook cnn ebay

nothingtoseehere.com

thisisnotabotnet.com

Page 10: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

H20  World,  11/10/2015  

   

     9  

FRAUD  DETECTION  

facebook cnn ebay

nothingtoseehere.com

thisisnotabotnet.com

Page 11: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

10  

•  Offline  data  processing  (Hadoop)  -  Batch  models:  flag  machines  and  sites  

•  Streaming  data  processing  (Storm  /  Kaea)  - Near  real-­‐Kme  online  models:  flag  machines  

•  In-­‐browser  measurement  -  Real-­‐Kme  rules  and  models:          flag  impressions  

 Combine  all  signals  in  real  Kme  to  flag  fraudulent  impressions  

H20  World,  11/10/2015  

FRAUD  DETECTION  COMBINATION  

Page 12: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

HUMAN  INTERACTION  

Page 13: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

H20  World,  11/10/2015  

         

12  

SOME  METRICS  MAY  BE  MISLEADING  MOUSE  HOVER  RATE  

Page 14: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

VIEWABILITY  

         

13  H20  World,  11/10/2015  

Page 15: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

•  How  many  Kmes  the  driver  sees  the  billboard?  •  How  ofen?  •  Is  there  enough  Kme  to  read  it?  

IS  VIEWABILITY  ALL  THAT  MATTERS?  

         

14  H20  World,  11/10/2015  

THE  IMPORTANCE  OF  THE  ENVIRONMENT  

Page 16: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

H20  World,  11/10/2015  

         

15  

Page 17: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

H20  World,  11/10/2015  

         

16  

Page 18: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

H20  World,  11/10/2015  

         

17  

HOW  CAN  ONE  TELL  IF  A  METRIC  MATTERS?  RANDOMIZED  A/B  TESTING  IS  GREAT  

Campaign Ad PSA

But sometimes…

Treatment Control

Page 19: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

NATURAL  EXPERIMENTS  

         

18  H20  World,  11/10/2015  

OBSERVATIONAL  CAUSAL  ESTIMATION  

versus

bit.ly/ias-­‐kdd15-­‐causal  

Page 20: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

CAUSAL  LIFT  VS  NUMBER  AND  LENGTH  OF  EXPOSURE  

         

19  H20  World,  11/10/2015  

RESULTS  FOR  A  TELECOM  CAMPAIGN  

1 45 9000.10.20.30.40.5

120756030155

N = 9 impressions

Caus

al lif

tExposure time (sec)

Caus

al lift

N impressions1 5 10

Exposure = 15-30 secs

00.10.20.30.40.5

Page 21: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

CAUSAL  LIFT  VS  NUMBER  AND  LENGTH  OF  EXPOSURE  

         

20  H20  World,  11/10/2015  

RESULTS  FOR  A  TELECOM  CAMPAIGN  

Page 22: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

21  

•  FighKng  ad  fraud  is  an  ongoing  challenge  -  Faster  detecKon  helps  

•  We  are  able  to  tell  which  metrics  are  meaningful  -  Causal  inference  is  criKcal  - Mouse  hover  rate  is  not  a  good  measure  of  engagement  -  Viewability  is  important,  but  is  not  the  ulKmate  goal  or  means  

- Number  and  frequency  of  exposure  mahers  -  Total  exposure  duraKon  mahers  - One  long  exposure  doesn’t  quite  reach  the  goal  

H20  World,  11/10/2015  

SUMMARY  

Page 23: H2O World - Learning How Humans and Non-Humans Interact with Digital Ads

THANK  YOU!  [email protected]