Machine Learning
Applications in Marketing
Day 5
Gokhan Yildirim
Outline
• Unobserved data and consumer privacy
• Case study: DineTogether
• How to turn better data into better
decisions?
• Module wrap up
Learning objectives
• Identify the potential issues related to privacy and value
of individual –level data to consumers and companies.
• Discuss the role of machine learning in marketing
decision making.
• Explain why some machine learning applications fail in
practice and suggest solutions.
• Summarise the key learnings from the module.
Unobserved data and
privacy
Amazon Shopper Panel
• The Amazon Shopper Panel (ASP) is
an opt-in, invitation-only program
• Participants can earn monthly rewards
by sharing receipts from purchases
made outside of Amazon
• Upload 10 eligible receipts per month
by using the ASP app and earn £5.
POLL
Is the Shopper Panel a good idea or a bad
idea for Amazon? Why?
❑ Good
❑ Bad
POLL
Does the Amazon Shopper Panel represent
an opportunity or a threat to consumers’
control of their privacy and data?
❑ Opportunity
❑ Threat
The more data, the
better?
Is more data always better?
Option 1
Option 2
Value from data
Option 3
Option 4
Amount of data
Is more data always better?
Option 1
Option 2
Value from data
Option 3
Option 4
Amount of data
Three dimensions of data
N
T
J
Case study: DineTogether
Which option(s) would you pick?
1. Anonymous auditions. Remove applicants’ profile pictures
to direct hosts’ attention to the candidates’ food experience
preference instead.
2. Total host transparency. Allow candidates to see hosts’
previous venues, menus, and guests. Let candidates
determine if an event is for them.
3. AI matchmaker. Remove the application process entirely
and use AI technology to match guests to hosts’ events.
4. Limit to three preferences out of nine total
options. Potential guests would specify three preferences
(down from nine) among the categories of food, venue, and
host, thereby reducing the pool of applicants for each event
and improving their chances of being accepted.
Simulation game
debrief
Consumer Products Industry
• Consumer products are a data-intensive industry, and
almost every company in the industry is becoming
increasingly analytical.
• Much of the industry’s customer data comes ultimately
from bar code scanners within supermarkets.
• Use of data and analysis can help executives in that
industry learn whether they are targeting the right
customers, pursuing the right strategies, operating
their business in the right way, and undertaking
16 the
right marketing campaigns and promotions.
How to turn better
data into better
decisions?
Wonderful marketing analytics for
today’s data-rich environments
But do they improve decisions ?
• “Our organization has more data than we
could possibly use” – A Global Brand Director
• “I have more data than ever, less staff than
ever, and more pressure to demonstrate
marketing impact than ever”—A CMO
Big data project issues
• > 55% of big data projects not completed;
even more fail to meet expectations (Iyer
2014)
• Big data passed Hype
Cycle, moves through
Trough of
Disillusionment
Big data is often…
“like sitting an exam and not bothering to read the
question,
simply writing out everything you know on the
subject and,
hoping it will include the information the examiner
is looking for.”
Big data should be…
“about the interface between the analytical,
experimental science that goes on in data labs,
and the profit and target chasing of sales force
in the boardroom”
When
Big Data
Goes
Bad
Examples of big data gone bad
• Google Flu Trends
(GFT) predicts winter
more than the flu:
– residual
autocorrelation +
seasonality
Why does this happen?
• Big data hubris: big data assumed to substitute
traditional data collection & analysis
– e.g. GFT underperforms other flu models but can be combined as it provides
complementary info
• Measurement dynamics : Google updates its
algorithm often for profits & ‘popular’ terms
makes search endogenous
Our take: human biases (3C’s) match the
3Vs of big data:
1. Confirmation bias:
1. Volume: Increasing With more data, you have more
amount of data opportunity to find confirmation for
your idea or your prior beliefs.
2. Communication bias:
2. Variety: Different Increased complexity of data, metrics
and analyses make it difficult to
types of data communicate insights for decision
making.
3. Velocity: Real time 3. Control Illusions:
Increased Fast changing, real-time
data metrics give illusion of control, but are
not leading KPIs
Source: Seggie, Soyer and Pauwels (201)
CHALLENGES
1. Volume challenge
DATA= PATTERN + NOISE
Build a model that separates noise from the
pattern well.
2) Variety challenges
• ‘My colleague in charge of social listening brings great
insights, but he can’t tell me why consumer said it and
in what context”
Barry Jennings, Global Marketing Insights Director, Dell
2) Variety challenges
• ‘A limitation of analytics which only make use of
customer records is that intangible but important
variables such as brand awareness, image and
attitudinal data, are absent’
Kevin Gray
Integrate slow moving attitudes and
fast online actions
DO
KNOW
Search LIKE
COGNITION Click AFFECT
Aware Visit Prefer
Web visits
Loyalty
Buy
Consider
Experience
& Express
Visualize & Simulate Simply: a slide bar
32
Heatmaps explore feasible
profit lifts
Heat Map of the Interaction of Two Marketing Variables on Profits
Price in $
#REF!
10 15 20 25 30 35 40 45 50 55 60 65 70 75
0
0.02 1.04 1.92 2.64 3.22 3.65 3.93 4.06 4.04 3.87 3.56 3.09 2.47 1.71
250 0.65 1.68 2.56 3.28 3.86 4.29 4.57 4.70 4.68 4.51 4.19 3.73 3.11 2.35
500 1.25 2.27 3.15 3.87 4.45 4.88 5.16 5.29 5.27 5.10 4.79 4.32 3.70 2.94
TV advertising in thousands of $
750 1.79 2.81 3.69 4.41 4.99 5.42 5.70 5.83 5.81 5.64 5.33 4.86 4.24 3.48
1000 2.28 3.30 4.18 4.91 5.48 5.91 6.19 6.32 6.30 6.13 5.82 5.35 4.73 3.97
1250 2.72 3.74 4.62 5.35 5.92 6.35 6.63 6.76 6.74 6.58 6.26 5.79 5.18 4.41
1500 3.11 4.13 5.01 5.74 6.32 6.74 7.02 7.15 7.13 6.97 6.65 6.18 5.57 4.80
1750 3.45 4.48 5.35 6.08 6.66 7.09 7.37 7.50 7.48 7.31 6.99 6.52 5.91 5.14
2000 3.74 4.77 5.65 6.37 6.95 7.38 7.66 7.79 7.77 7.60 7.28 6.82 6.20 5.44
2250 3.99 5.01 5.89 6.62 7.19 7.62 7.90 8.03 8.01 7.84 7.53 7.06 6.44 5.68
2500 4.18 5.21 6.08 6.81 7.39 7.81 8.09 8.22 8.21 8.04 7.72 7.25 6.64 5.87
2750 4.32 5.35 6.23 6.95 7.53 7.96 8.24 8.37 8.35 8.18 7.86 7.40 6.78 6.02
3000 4.42 5.44 6.32 7.05 7.62 8.05 8.33 8.46 8.44 8.27 7.96 7.49 6.88 6.11
3250 4.46 5.49 6.36 7.09 7.67 8.10 8.38 8.51 8.49 8.32 8.00 7.54 6.92 6.15
3500 4.46 5.48 6.36 7.09 7.66 8.09 8.37 8.50 8.48 8.31 8.00 7.53 6.91 6.15
3750 4.40 5.43 6.30 7.03 7.61 8.04 8.32 8.45 8.43 8.26 7.94 7.48 6.86 6.09
4000 4.30 5.32 6.20 6.93 7.50 7.93 8.21 8.34 8.32 8.15 7.84 7.37 6.75 5.99
4250 4.14 5.17 6.04 6.77 7.35 7.78 8.06 8.19 8.17 8.00 7.68 7.22 6.60 5.84
4500 3.94 4.97 5.84 6.57 7.15 7.57 7.85 7.98 7.97 7.80 7.48 7.01 6.40 5.63
4750 3.69 4.71 5.59 6.32 6.89 7.32 7.60 7.73 7.71 7.54 7.23 6.76 6.14 5.38
5000 3.38 4.41 5.29 6.01 6.59 7.02 7.30 7.43 7.41 7.24 6.92 6.46 5.84 5.08
Visualize effectiveness over time
34
Compare profit from different
scenarios
| 35
How to turn Data into Decisions ?
Big Data Challenges My Advice
V’s C’s
Volume Confirmation Identify & test
hypotheses fast
Variety Communication Visualize &
simulate with the
right metrics
Velocity Control Loop in Build-
Measure-Learn
Source: Seggie, Soyer and Pauwels (2017)
Why ‘traditional’ skills are key
• The biggest reason that investments in big
data fail to pay off is that most companies
don’t do a good job with the information they
already have.
• They don’t know how to manage it, analyze it
in ways that enhance their understanding,
and then make changes in response to new
insights.
Technology, Power,
and Prosperity
Technology
should be a tide
that lifts all boats
Power and Progress
• Innovation has often empowered a tiny elite while leaving the
majority of workers with little of the promised prosperity.
• Technology to be beneficial to society and create no
unemployment, inequality, or poverty, it needs to be
controlled by social power and government regulation.
• It should be about promoting transparency, treating
consumers fairly, and re-evaluating who controls the
means of information in our increasingly digital world.
• Progress is not automatic but dependent on the choices we
make about technology.
Module takeaways
Takeaways
• Making decisions that involve multiple variables can be
a complex balancing act.
• Marketing is a blend of art and science.
• There is good and bad marketing analytics
• Important to know under which conditions “science”
should rule and under which conditions “art” should rule.
What is the SCIENCE side?
Some of the successful applications are:
• Segmentation by observing and clustering the data
• Targeting the right consumers with the right message at the right
time
• Accessing and assessing social media and ratings information
• Which sales leads are worth pursuing
• Sales forecasts when preferences are stable
• Setting prices dynamically
• Sales-force compensation
• Optimization of ad spending
• Customer retention management
What is the ART side of SCIENCE?
Part of the art is:
• choosing the problems to look at,
• how to approach the problem,
• select the data,
• how to use and analyse the data,
• which variables to track and use for prediction,
• how to interpret the results, and how to utilize the
results.
Science allows for more informed “arts” decisions.
Quantitative intuition
Effective decision-making calls
for human intuition to join forces
with data instead of replacing it.
Machine learning application
fails if…
• You do not define the business problem clearly
• The model and data do not map to the business problem
• Data collected by different systems is disjointed
• You neglect data quality (e.g. unobserved factors, offline events) and
data sampling issues (e.g. representativeness of social media data)
• You use inappropriate tools and/or do not apply the model correctly
• You focus on performance metrics that do not matter
• You do not have the right talent to leverage machine learning.
Mela and Moorman (2018)
Your learning
Your feedback
• Please fill out the MODES survey
• Check your email now for the survey link
(hosting@[Link]) or use the QR
code here.
Your feedback
Please write down 3 things you learned in this
module on a piece of paper and hand it to me.
Module wrap-up
Thank you!
