a practitioner’s introduction to attribution...•rule-based attribution models do not fully value...
TRANSCRIPT
#ARFAxS
A Practitioner’s Introduction to Attribution
Mericcan Usta
Sr. Data Scientist
GroupM
• ISOLATE EFFECTS OF MEDIA
•OPTIMIZE ADSPEND MIX
A Working Definition ofAttribution
Legacy: Top-Down Models
• Relate aggregate media spend to aggregate sales• How do different media perform relative to each other? • Which marketing-mix modeling choices will help isolate the impact?
• Halo effects, adstock, short-term response curves, long-term effects, mixed effects modeling…
SOURCE: Kihn, M., C. Eubanks, and L. F. Kune, Market Guide for Attribution and Marketing Mix Modeling. Gartner Research Article ID G00317823. 30 May 2018.
A Paradigm Shift
• Promise of a universal single-source media panel
• Promise of a universal single-source panel
• Customer journey is the new currency
Now: Bottom-Up Models
• Find the critical media that keeps incremental journeys in the funnel• Is my data quality good enough?
• Did people really see my ad?• Can I isolate customer journeys?• How will data privacy rules affect the quality?
SOURCE: Kihn, M., C. Eubanks, and L. F. Kune, Market Guide for Attribution and Marketing Mix Modeling. Gartner Research Article ID G00317823. 30 May 2018.
New Media Imply New Models?
display across walled gardens
video platform genres
programmatic linear TV
IoT connections to OOH
“new media” reported in aggregate
“old media” amenable to journeys
A Taxonomy of Bottom-Up Models
Rul
e-ba
sed
Met
hods
Sta
tistic
s-ba
sed
Met
hods
Rule-based methods rely on business judgement and technical simplicity…
…statistics-based methods use data.
Findings from the JAR
• Heuristics are good, but some are too simple
• “Last click” model favors paid search, but consumers typically search multiple times before making a purchase
• Rule-based attribution models do not fully value social media or email, which often don’t directly lead to purchase but can have strong behavioral impact
• Statistics-based model provides more stable attributions
SOURCE: Nisar, T.M., M. Yeung. Attribution modeling in digital advertising. Journal of Advertising Research Dec 2018, 58 (4) 399-413
Rule-based
Statistics-based
Statistics–Based Models are AboutMarginal Value
CONVERSIONShapley value concept
Shapley Value Concept: An Example
Display
SearchSocial
• Play an “advertising game”• Players: publisher types where ads are placed• How much contribution to the advertiser’s bottom line does each “player” make
in each journey?
Shapley Value Concept: An Example
DisplaySearch
Social
Marginal contributions of each player (publisher types where ads are placed) averaged across all possible journeys
Solving the Scalability Challenge
SHAPLEY VALUE APPROXIMATIONSexponential complexity with budget alternatives logistic regression & probabilistic models
memory runtime
quality/cost balance
DIVIDE & CONQUERfeatures (minimal overlap) and journeys
Putting the Ideas to Test
• Campaign run for a nationwide accessories retailer between August 7th, 2016 to October 30th, 2016 through the retailer’s website
• 400M impressions served to 120M users, 1M associated with revenue generating customer journeys
• 137K online purchases
SOURCE: Mahboobi, S.H., M. Usta, S. R. Bagheri. Coalition Game Theory In Attribution Modeling. Journal of Advertising Research Dec 2018, 58 (4) 414-422
• ISOLATE EFFECTS OF MEDIA •DIGITAL PROMISES CUSTOMER JOURNEY
•NEWEST MODELS ATTEMPT MARGINAL CONTRIBUTION
•APPROXIMATIONS MAKE IT SCALABLE• CAVEATS
Recap
DATA SCIENCE
[email protected]: mcusta