Keynote of the 4th #DDTT meetup (http://digitaldata.tips) in Amsterdam, the Netherlands. Guy Liberali of RSM / Erasmus presented about optimization through algorithms: bandits and morphing websites.
Stay up to date about new #DDTT events through our meetup group: http://meetup.com/onlineoptimizers
Organizer: Webanalisten.nl / Online Optimizers meetup group
Sponsors: Adobe and Relay42
Gui Liberali, Morphing Websites - Digital Data Tips Tuesday #4 - Growth through Data - Amsterdam - May 19th 2015 - #DDTT
1. Morphing Theory and Applications
to Online Experiments:
Learning While Earning
DDTT 2015
Gui Liberali
Erasmus University
Laboratory for
Optimization of Digital
Experiments @ Erasmus
3. One size does not fit all
(a) General content, large-load, graphical morph (b) Focused, small-load, verbal morph
Visual
Technical
Content
More
Content
Audio
Less
Content
General
Content
Website designs as Morphs
Appeals to Verbal
Cognitive Style
Appeals to Visual
Cognitive Style
4. Appeals to Holistic, impulsive cognitive
style
Appeals to Analytical, deliberative
cognitive style
One size does not fit all
Banners as Morphs
5. • From learning-then-earning to the world of
learning-while-earning
• Higher CTR, conversions
• Better user experience
Where’s your A/B
Testing ?
• From aggregate to disaggregate data
• From demographics to cognitive styles
6. 1. Learning about Styles vs Morphs
Firm
Side:
Banner
Ads
Buy!
versus
Learn
More
Inform
a4ve
Emo4o
nal
Time-‐
pressu
re
Consumer
Side:
4
Cogni4ve
Styles
Verbal-‐Impulsive
0.1
0.11
0.2
0.13
Visual-‐Impulsive
0.11
0.2
0.13
0.1
Verbal-‐Delibera4ve
0.2
0.13
0.1
0.11
Visual-‐Delibera4ve
0.11
0.2
0.13
0.1
7. Optimization: from Earn to Learn
• Maximize conversion by picking the best among ads for current user
• If we knew the conversion probabilities perfectly we would simply serve the
optimal morph
• If we knew nothing about the conversion probabilities we would randomize
and learn
Need to balance learning
the true cognitive style
(exploitation) and providing
the optimal morph given
what we know now
(exploration).
For each person!
8. Morphing: real-time adaptation of the
firm’s actions to each consumer
Cognitive-style, segment, position in
the purchase funnel, ..
Website design, banner, advertising
campaign, media copy per channel,
promotion, price, product bundle, product
recommendation …
Consumer
latent
variable
Marketing
instrument
to be adapted
(morphed)
x
9. Morphing - Evidences from the Field
Morphing Banner Advertising (2014)
• Morphing almost doubled click-through rates for context-
matched banners (83%, p<0.01 and 97%, p=0.028) lifts,
respectively, for banners and for consumers
• Was live on CNET.com (8.4M visitors/day) for a month
• Extends behavioral targeting to avoid local max solutions
Website Morphing (2009)
• Online sales of BT group’s broadband plans can be
increased up to 20% by adapting the website to
match cognitive styles using our adaptive learning
algorithm
Website Morphing 2.0 (2014)
• NPV reward improvement of 69% over the NPV of
the original algorithm; corresponds to $17.5 million
for BT group if implemented
• Proof-of-Feasibility on Suruga Bank
• Generalized version of morphing, including when to
morph, switching costs, multiple morphs
10. Morphing Algorithm
Challenges Addressed
1. What is the style of the current user?
– Update our beliefs as new clicks become available
2. What is the optimal morph for each style?
– Update as we observe successes and failures
11. Morphing
Choice
descriptor
Priors
on
distribu5on
of
styles
Updated
user
style
descriptor
On-‐line
Op4miza4on
Pgm:
probability
of
purchase
for
individual
in
style
g
given
morph
m
Banner
Ads
-‐>
1
2
3
4
Cogni4ve
Style
Impulsive
Delibera4ve
0.1
0.11
0.2
0.13
Illustra5ve
example
User
Op5mal
morph
for
current
user
Website
User
makes
a
purchase
or
quit
Bayesian
update
12. Optimal solution with Gittins’ indices
(assume we know cognitive style)
Gittins’ indices for the
eight morphs.
Morph that
was chosen.
0
1
2
3
4
5
6
7
8
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Visitor
ChosenMorph
0.35
0.45
0.55
0.65
0.75
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Visitor
Gittins'Indices
Morph 0 Morph 1 Morph 2 Morph 3 Morph 4 Morph 5 Morph 6 Morph 7
System experiments with Morph 3
for a while before settling back to
Morph 2.
13. • What is the cognitive style of the current user?
– Update our beliefs as new clicks become available
• What is the optimal morph for each style?
– Update as we observe successes and failures
Website
Morphing
(2009)
• How to handle repeat visits and cumulative learning ?
– Field experiment on cnet.com w/ AT&T banner ads
– Roll-back mechanism
Morphing
Banner
Advertising
(2014)
• Generalized Morphing Algorithm: w/ When to Morph
– DP solves trade-off between gains from morphing early (longer
exposure) and gains from morphing later (better knowledge)
– Exposure to multiple morphs
• Attribution problem: recency, exposure, both
• Switching costs: changing morphs only if expected rewards exceed
penalty for switching
Website
Morphing 2.0
(2014)
Morphing Trilogy
14. Morphing in the Field
• Can be used normatively (e.g., field experiments) or descriptively (secondary
data)
• Two tuning parameters makes it flexible to various applications
– Switching costs
– Different attribution theories/regimes
• Morphing is appropriate to contexts with
– High volume of observations
– Clear success/failure outcome
– Latent variable
15. Laboratory on Optimization of Digital
Experiments (LODE) @ Erasmus University
• Executive education course on Digital Experimentation and Big Data
• Research Fridays: Bi-monthly meetings of academics and founding companies at
Erasmus University
• Bi-yearly publication of research priorities for research on A/B testing
• Erasmus students running A/B testing & morphing with real-world company traffic
– Module 1: Text-analysis methods
– Module 2: A/B experiments
– Module 3: Morphing
16. References
• Hauser JR, Urban GL, Liberali G, Braun M (2009) Website morphing. Marketing
Sci. 28(2):202–224.
• Urban G., Liberali G., MacDonald E., Bordley R., Hauser J. (2014) Morphing
Banner Advertising, Marketing Science, 33(1): 27-46.
• Hauser J., Liberali G., Urban G. (Summer 2014) Website Morphing 2.0: Switching
Costs, Partial Exposure, Random Exit, and When to Morph. Management Science.