2. Marketing departments are more important than ever to serve the ambitions of
organisations.
NG LEADERS MUST DEMONSTRATE THE RETURNS THEY
3. After 3 years of consecutive growth,
marketing budget
growth
has stalled
67%
of CMOs plan to increase
their spending in Digital
advertising
Numbers and Insights - Taking action
4. Marketing executives are required to track their spending.
Transforming the business
from this … to that
5. Marketing attribution: transforming a business question into a data science
problem
• How can I quantify the
influence an advertisement
has on a customer decision to
make a purchase?
• How can we measure the
effectiveness of each channel?
7. Marketing attribution: heuristic modelling
• Arbitrarily decide to reward on
segment of the journey
Last click ?
Major problem: will introduce a bias
towards retargeting. What if my customers
are not that sensitive to retargeting but
more to a good old Google search?
First click ?
Major problem: will introduce a bias
towards Google search. What if my
customers are sensitive to social media ?
8. Marketing attribution: algorithmic modelling
• All rule-based models
introduce some human bias
whatever the « complexity » .
• How can we use data
science to solve that
problem?
9. Introducing deep recurrent artificially
convoluted networks for distributed machine
learning models in marketing attribution
• How does that sound? Cool? Smart? Trendy?
• Data science solves well-framed problems in a well-framed manner.
• Weird…? Too complicated? Indeed. And it should not.
10. Attribution Modelling = Track your channels + Record the conversions
1. Markov Chain Modeling
3. Game theory and Shapley value
2. How to validate the best approach ?
Define and validate a model
11. A good data science solution starts … with good data
Prerequisites
• Track all your user
actions on each
targeted channel
• Leverage all the
progress IT has done
recently.
• And keep in mind that
there will certainly be
some limitations and
missing data
12. Marketing attribution: Markov Chain modelling
The methodology: from raw log files to a Markov Model
• From user actions database, compute each
customer journey
• From each customer journey, compute the
Markov transition matrix of a typical
customer
• From the Markov transition matrix, compute
the standard likelihood to convert
• From the standard likelihood to convert,
compute the removal effect of each
channel.
• Really hard to do A/B
testing in practice
• Too expensive and risky
to stop a channel
• Sessionalization in database (30
minutes windows)
• Distributed aggregation with
Spark
13. Marketing attribution: Markov Chain modelling
C1
(start)
C2 C3
(Null)
(Conv)
Computing the standard likelihood to convert
66.7%
33.3%
50%
100%
50%
50%
Let’s compute P(conv) which is the likelihood of any given customer to
convert.
50%
14. Marketing attribution: Monte Carlo Markov Chain
C1
(start)
C2 C3
(Null)
(Conv)
Computing the removal effect of a given channel
66.7%
33.3%
100%
50%
50%
Let’s compute how the likelihood of conversion will be affected when we
remove a channel P(conv_without_C1)
Removal effect = P(conv_standard) - P(conv_without_C1)/P(conv_standard)
16. Marketing attribution: Markov Chain modelling
Each channel has a removal effect; we can
now
• Compare them over time
• Spend the budget based on the actual
effect of each one
Easy, robust and fast technique
• can easily be deployed in production
• lead to quick win in the organisation
Summary
Currently some live implementations and
on our own data and some of our clients.
Promising results :)
By the way, here is a non-exhaustive list of
our happy customers (where we might have tested this
approach)
Consumer Goods E-Retail
Travel & HospitalityServices
18. How much value does a
channel bring to an already
existing group of channels ?
Marketing attribution: Game theory and Shapley Value
Facebook + Twitter = 2% of
conversions.
What if we add SnapChat
down the line ?
2 %
3 %
Increased likelihood
of purchase by 50%
Likelihood
of purchase
Customer journey
leading to conversion
19. Marketing attribution: Game theory and Shapley Value
Methodology and mathematical reasoning
• From user actions database, compute
each customer journey and the
number of conversions.
• Compare the added value of a
channel on the number of conversions
to every coalitions channels.
• Normalize data: study % of
conversions and not number
conversions
• Scientific approach: keep in
mind some basic reasoning
like symmetry
Work in database.
Use SparkSql and
PySpark
20. Marketing attribution: Game theory and Shapley Value
{Facebook} = 5 conversions
{Twitter} = 2 conversions
{SnapChat} = 3 conversions
{Facebook , Twitter} = 8 conversions
{Facebook , SnapChat} = 13
conversions
{Twitter , SnapChat} = 7 conversions
{Facebook , Twitter , SnapChat} = 16
conversions
Facebook contribution
(5 + 6 + 10 + 9)/ 4 = 7.5
Twitter contribution
(2 + 3 + 4 + 5) /4 = 3.5
SnapChat contribution
(3 + 8 + 5 + 8 )/4 = 6
A practical example from the different customer journey leading to conversion
22. Marketing attribution: model validation strategy
Practical difficulties of the real world
How do you observe the real removal
impact? Quantify the real contribution of a channel?
Hard to verify the grand truth of what you
can compute fro the database.
A/B testing
• Potential disastrous effect for the business to stop
one channel even for a day.
• Client acquisition is more challenging than ever.
Why should you do that ?
• To choose the best approach
• To monitor the potential drift in real time
23. Marketing attribution: model validation strategy
Methodology for validation strategy
Random simulation
of customers journey
Apply attribution models
Define conversions
given an hypothesis H
How did models
capture H?
Deploy the best model
to allocate
marketing budget
Increase your
conversion rate
24. Marketing attribution: model validation strategy
Results on Dataiku logs
Hypothesis: noisy data
Shapley can capture that
much better than any other
(No money should be
spent there)
Caution!!
Be careful that there is
a natural bias in the data
Logs in database are
probably … the results of
actual marketing campaign
25. A DSS workflow eases the process of
iterative data science
• Handle the mandatory data
preparation
steps
• Handle some complex coding
transformation in SQL/Python/R/Scala
• Leverage the underlying infrastructure
Dataiku DSS the most complete Data Science Platform
A DSS workflow offers some features to
easily deploy in production
• Scheduling and automation of jobs
• Advance usage of the internal API
• Deployment without recoding anything