Human Factors of XR: Using Human Factors to Design XR Systems
Machine Learning driven Quantum Optimization for Marketing
1. www.deeplearningitalia.com 1
DAL DEEP LEARNING AL QUANTUM COMPUTING
Le nuove frontiere tecnologiche che stanno cambiando il nostro presente e il futuro
In collaborazione con: Fondazione Bruno Kessler
2. Machine Learning driven Quantum
Optimization for Marketing
Calogero Zarbo
Senior AI Research Scientist @Satispay
www.deeplearningitalia.com
3. 3
The Problem
Our main Consumer acquisition channel is the MemberGetMember (MGM) campaign,
counting for more than 40% of the total consumer acquisition.
MGM campaigns are a substantial % of the overall marketing budget: about 25M Euros in
2022.
Even slight optimizations in this area may bring in huge results in terms of savings
www.deeplearningitalia.com
4. 4
The Solution
Personalized MGM campaigns
The assumptions are:
● Each user has unique pattern of being engaged with the app
● These patterns can be proxied by how the app is being used
Why you need AI:
● Not all customers sees every possible incentives a reasonable number of times
● Fill the gap with real measured data is costly
www.deeplearningitalia.com
5. 5
Data Pipeline & Strategy
Interaction with the BI team is crucial to make all of this happens
● Extract key information from internal datalake about registration, activation,
transactions and MGM activity
● Ground preparation for A/B test (Power Analysis)
BI not only provide data, but strategize the whole testing process:
1. Define the number of consumer we can work with
2. Stratified sampling at district level for both test and control group
3. Handle the communication with marketing for the chosen incentive allocation
4. Results harvesting and KPI evaluation
www.deeplearningitalia.com
6. 6
The Target
We experiment various type of target, but we decided that the most important ones
would be:
● Mean of the #Matches @ incentive level (appreciation)
● Presence of the Match @ incentive level (activation)
They express two slightly different perspectives in the user appreciation of a particular
campaign.
www.deeplearningitalia.com
8. 8
The Models
We tried different models:
● LightGBM
● XGBoost
● Transformers
After setting up a proper Data Analysis Plan (DAP), we discarded the first two options due
to their weakness towards unbalanced datasets.
Matthew Correlation Coefficient (MCC) < 0.4
www.deeplearningitalia.com
13. 13
The Setup so far..
We choose a time span of 6 months history
For each Incentive we trained:
● A transformer for the Activation flag in the last 6 months (classification)
● A transformer for the Mean #invitation in the last 6 months (regression)
Phase 1 - The Backtesting
● Activation flag: MCC: 0.91 overall → Good to go
● Mean #invitation: R2 Score: 0.31 overall → Good to go
Phase 2 - The A/B Test
● Decision making process missing
www.deeplearningitalia.com
15. 15
The decision maker
Now that I have a level of appreciation for each user, for each incentive level, and a
probability of activation for each user, for each incentive level how to decide who gets
what incentive?
www.deeplearningitalia.com
16. 16
The decision maker
Now that I have a level of appreciation for each user, for each incentive level, and a
probability of activation for each user, for each incentive level how to decide who gets
what incentive?
All possible combinations of (User,Incentives)?
www.deeplearningitalia.com
17. 17
The decision maker
Now that I have a level of appreciation for each user, for each incentive level, and a
probability of activation for each user, for each incentive level how to decide who gets
what incentive?
All possible combinations of (User,Incentives)?
This seems like an NP problem: which allocation would maximize the appreciation keeping
the cost under a certain target?
www.deeplearningitalia.com
18. 18
Mathematical formalization
pij
= predicted appreciation of ith
user on jth
incentive
xij
= binary variable 0/1 whether or not assign that particular ith
user the jth
incentive
cj
= the cost of the jth
incentive
aij
= predicted activation of ith
user on jth
incentive
max sumij
(xij
* aij
* pij
)
s.t.
sumi
(xij
) = 1 for each jth
user
sumij
( (xij
* aij
* pij
* cj
) - CAC target) = 0
sumij
(xij
* aij
* pij
) >= Appreciation target
www.deeplearningitalia.com
22. 22
D-Wave Quantum Annealer
www.deeplearningitalia.com
Entanglement:
Two QuBits interact with each other without being
directly stimulated.
Tunnelling:
A QuBit can go in any better place in the solution space
without paying the price.
Superposition:
A QuBit can be 0 and 1 at the same time
28. 28
The first A/B Test
www.deeplearningitalia.com
Number of People: 16k for each group
Incentives: 20€, 30€, 40€, 50€
Geographic area: all north-west of italy
Sampling method: 1% of each district
Can the system trigger more invitations
keeping the cost close to the group A?
29. 29
The first A/B Test
www.deeplearningitalia.com
TARGET
MISSED
Number of People: 16k for each group
Incentives: 20€, 30€, 40€, 50€
Geographic area: all north-west of italy
Sampling method: 1% of each district
Can the system trigger more invitations
keeping the cost close to the group A?
- No
Cost - p-value: 0.0188
Match - p-value: 0.071
30. 30
What was wrong?
● Is the whole system not suitable for the task?
● Were the training data not representative?
● Did we introduce any confounding factor?
www.deeplearningitalia.com
31. 31
What was wrong?
● Is the whole system not suitable for the task?
● Were the training data not representative?
● Did we introduce any confounding factor?
Our speculations were:
1. The appreciation variables already include the activation information. Provide both of
them might be misleading. Let’s remove it.
2. The mean of # of matches computed on 6 months might not be representative
enough. Let’s try 3 months.
www.deeplearningitalia.com
32. 32
The whole system -V2
www.deeplearningitalia.com
BI
Appreciation
Activation
Decision
Making
Machine
Marketing
33. 33
Mathematical formalizationV2
pij
= predicted appreciation of ith
user on jth
incentive
xij
= binary variable 0/1 whether or not assign that particular ith
user the jth
incentive
cj
= the cost of the jth
incentive
aij
= predicted activation of ith
user on jth
incentive
max sumij
(xij
* aij
* pij
)
s.t.
sumi
(xij
) = 1 for each jth
user
sumij
( (xij
* aij
* pij
* cj
) - CAC target) = 0
sumij
(xij
* aij
* pij
) >= Appreciation target
www.deeplearningitalia.com
34. 34
Mathematical formalizationV2
pij
= predicted appreciation of ith
user on jth
incentive
xij
= binary variable 0/1 whether or not assign that particular ith
user the jth
incentive
cj
= the cost of the jth
incentive
max sumij
(xij
* pij
)
s.t.
sumi
(xij
) = 1 for each jth
user
sumij
( (xij
* pij
* cj
) - CAC target) = 0
sumij
(xij
* pij
) >= Appreciation target
www.deeplearningitalia.com
35. 35
The second A/B Test
www.deeplearningitalia.com
Number of People: 25k for each group
Incentives: 20€, 30€, 40€, 50€
Geographic area: all north-west of italy
Sampling method: 1.5% of each district
Can the system trigger more invitations
keeping the cost close to the group A?
36. 36
The second A/B Test
www.deeplearningitalia.com
Number of People: 25k for each group
Incentives: 20€, 30€, 40€, 50€
Geographic area: all north-west of italy
Sampling method: 1.5% of each district
Can the system trigger more invitations
keeping the cost close to the group A?
- It seems so
TARGET
HIT
Cost - p-value: 0.0125
Match - p-value: 1.56e-9
37. 37
The third A/B Test
www.deeplearningitalia.com
Number of People: 25k for each group
Incentives: 20€, 30€, 40€, 50€
Geographic area: all north-west of italy
Sampling method: 1.5% of each district
Can the system trigger more invitations
decreasing also the cost?
38. 38
The third A/B Test
www.deeplearningitalia.com
Number of People: 25k for each group
Incentives: 20€, 30€, 40€, 50€
Geographic area: all north-west of italy
Sampling method: 1.5% of each district
Can the system trigger more invitations
decreasing also the cost?
- We can’t say
TARGET
UNK
Cost - p-value: 0.0712
Match - p-value: 0.0573
39. 39
Cost Analysis - The Campaign
To maximize the campaign returns we need to run simulations at different CAC target as
follows:
● Start: 30€, End: 50€, Step: 0.5€
● Total of: 40 simulations per campaign
● For this exercise let’s optimize a campaign of about 50 000 users
● Total number of variables per campaign: #Users x #Incentives (20€, 30€, 40€, 50€) =
200 000 variables
● Campaign frequency: roughly one campaign every 2 weeks
www.deeplearningitalia.com
40. 40
Cost Analysis - HiGHS
● Time per simulation: ~37 hours
● Machine type: c7g.8xlarge
● Machine cost: 1.15$/h - 20% discount: 0.92 $/h
● Cost per simulation: ~34$
● Cost per campaign: ~1361$ -> 62 computing days -> unfeasible
NB. Reducing the number of simulations, reduce dramatically the quality of the tuned
campaign.
www.deeplearningitalia.com
41. 41
D-Wave Quantum Solver
● Time per simulation: ~5 minutes
● Machine type:Advantage 6.2
● Machine cost: 125$/h - no discount - we will negotiate it eventually
● Cost per simulation: 11$
● Cost per campaign: 440$ -> 3.3 computing hours -> feasible
NB.The cost is computed assuming the solver is used only for MGM campaigns, which is
unlikely since we will have other optimization projects that would leverage such
technology and therefore lower the offset price. (i.e. optimal top-up for fraud exposure
minimization)
www.deeplearningitalia.com
43. 43
Final* Remark
AI do not market your product.AI do not create engagement.AI do not make satispay
more appealing.
AI use data to point out where the great work of marketing is having a tangible positive
impact.
Without the talent of our colleagues, in all departments, there is no AI that can make
Satispay what is now and what will be in the future.
www.deeplearningitalia.com
45. 45
45
SPAZIO ALLE DOMANDE!
www.deeplearningitalia.com
In collaborazione con: Fondazione Bruno Kessler
DAL DEEP LEARNING AL QUANTUM COMPUTING
Le nuove frontiere tecnologiche che stanno cambiando il nostro presente e il futuro