LEARNING HOW HUMANS AND NONHUMANS INTERACT WITH DIGITAL ADS H2O World, November 10, 2015 Sergei Izrailev SVP, Data Science [email protected]
LEARNING HOW HUMANS AND NON-‐HUMANS INTERACT WITH DIGITAL ADS
H2O World, November 10, 2015
Sergei Izrailev SVP, Data Science [email protected]
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
2
DIGITAL ADVERTISING
MARKETING MESSAGE “Right Offer”
1 AUDIENCE TARGETING “Right Person” 2
MEDIA PLACEMENT “Right Place” 3
CAUSAL IMPACT & USER ENGAGEMENT
“ROI”
• 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
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
H20 World, 11/10/2015
5
DATA COLLECTION AND PROCESSING STACK
NON-‐HUMANS: AD FRAUD
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
H20 World, 11/10/2015
8
FRAUD DETECTION
facebook cnn ebay
nothingtoseehere.com
thisisnotabotnet.com
H20 World, 11/10/2015
9
FRAUD DETECTION
facebook cnn ebay
nothingtoseehere.com
thisisnotabotnet.com
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
HUMAN INTERACTION
H20 World, 11/10/2015
12
SOME METRICS MAY BE MISLEADING MOUSE HOVER RATE
VIEWABILITY
13 H20 World, 11/10/2015
• 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
H20 World, 11/10/2015
15
H20 World, 11/10/2015
16
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
NATURAL EXPERIMENTS
18 H20 World, 11/10/2015
OBSERVATIONAL CAUSAL ESTIMATION
versus
bit.ly/ias-‐kdd15-‐causal
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
CAUSAL LIFT VS NUMBER AND LENGTH OF EXPOSURE
20 H20 World, 11/10/2015
RESULTS FOR A TELECOM CAMPAIGN
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
THANK YOU! [email protected]