Jakub Štěch na konferenci Data Restart 2020: Demonstrujeme propojení online a offline dat na případové studii České spořitelny. Jak vytvořit zákaznickou 360° a kde jsou možná omezení? Jak s cílením souvisí NLP a neuronové sítě? Proč je důležité mít celé datové prostředí u sebe jako klient?

1. Jakub Stech
Data Science architect in DataSentics
Responsibilities:
• Find & Translate business problems across
projects for data science teams (Ceska Sporitelna)
• Personalizing user experience using data and
machine learning approaches
• Help building data science products (CV, NLP, …)

2. © 2019 Adpicker, a product developed by European Data Science Center of Excellence
DataSentics in a nutshell
How We Work Machine learning
30+
Data engineering
Cloud
25+
ML Center of Excellence
in Prague & Brno
Partners of
Key References
"Make data science and machine learning have a real impact on
organizations across the world...
...bring to life transparent production-level data science.
Quality inspection
Computer-vision-based AI solution for
automatically analyzing the quality of
placement of my products in a store shelf
Customer Engagement 360°
Combining digital and offline behavioral data
and machine learning to personalize the
customer experience
Fraud / AML / Risk
Using advanced methods such
as NLP and graph analytics to
identify anomalies, etc.
Contact center
Using NLP for automatic
routing of emails/request,
topic detection, etc.

3. AGENDA
• 360 in CSAS
• Case study I: personalised CRM campaigns
• Case study II: contextual online targeting

4. Leading every client to prosperity
=
Data-driven advisory based on
clients needs and real-time situations
4

5. … is not easy
Customer-
centricity….
Low frequency of
measured interactions
between a client in
offline channels

6. 100 Things,
We touch our
phones 2,617 times
a day, says studyAround 100
sessions every
day…

7. Typical internal/CRM data
Age/sex/address
Contracts
Sales channels
…

8. Digital „footprints“
Ad interactions
Web interaction
Apps
…

9. Typical CRM data Digital „footprints“
• Static, mostly long-term behaviour
• Facts and transactions
• Structured
• Easy to process with traditional tech
• Dynamically changing, reflecting needs
• Uncertainty, fragments of interests, lifestyle
• Enormous data
• Messy, unstructured, changing interfaces

10. Emailing /
SMS / Push
Branches
& sales
networks
Transactional
data
Callcenter
data / call logs
Classic
CRM / data
processes
Classic campaign
management tools
ONPREM
Client profile (CRM)

11. Client profile (CRM)
Classic campaign
management tools
Emailing /
SMS / Push
Branches
& sales
networks
Transactional
data
Callcenter
data / call logs
Classic
CRM / data
processes
Digital
channels
behaviour
INTERNET BANKING
CONNECT ONLINE AND OFFLINE DATA

12. Client profile (CRM)
Classic campaign
management tools
Emailing /
SMS / Push
Branches
& sales
networks
Transactional
data
Callcenter
data / call logs
Classic
CRM / data
processes
INTERNET BANKING
Digital
channels
behaviour
Digital campaign
management tools
Digital non-client & client behaviour profiles
Machine
learning
CREATE DATAPLATFORM IN CLOUD

13. Client profile (CRM)
Classic campaign
management tools
Emailing /
SMS / Push
Branches
& sales
networks
Transactional
data
Callcenter
data / call logs
Classic
CRM / data
processes
Digital
channels
behaviour
ONLINE ADVERTISING INTERNET BANKING
ADDING ONLINE CHANNELS: CASE BY CASE, TRANSPARENT VENDOR
Digital non-client & client behaviour profiles
Machine
learning
Digital campaign
management tools

14. Client profile (CRM)
Classic campaign
management tools
Emailing /
SMS / Push
Branches
& sales
networks
Transactional
data
Callcenter
data / call logs
Classic
CRM / data
processes
Digital
channels
behaviour
ONLINE ADVERTISING WEBSITE INTERNET BANKING
ADDING ONLINE CHANNELS: TRACKING
Digital non-client & client behaviour profiles
Machine
learning
Digital campaign
management tools

15. Omnichannel Banking Experience
Case study #2
15

16. Client profile (CRM)
Classic campaign
management tools
Emailing /
SMS / Push
Branches
& sales
networks
Transactional
data
Callcenter
data / call logs
Classic
CRM / data
processes
ONLINE ADVERTISING WEBSITE APP INTERNET BANKING
Digital
channels
behaviour
Digital campaign
management tools
Digital non-client & client behaviour profiles
Machine
learning
AI-augmented Customer Engagement 360°

17. ADSERVER
DSP
…
MASTER DATA
Automated download
CLOUD
PLATFORM
STORAGE
HOW OUR AI REALLY WORKS?

18. AUTOMATICALLY CREATING USER JOURNEYS

19. ALGORITHM
(optimizing
towards goal)
…
MASTER DATA
CLOUD
PLATFORM
STORAGE
(user journeys)
HOW OUR AI REALLY WORKS?

20. Mortgage / Loan / Saving account / Investments …
What product?
DIGI DATA BASED SALES SIGNALS

21. MACHINE LEARNING WHAT DRIVES THE GOAL
USER PATH TARGET
1. IMPRESSION 2. IMPRESSION 3. IMPRESSION LOAN
finance sauto sauto 1
finance sreality - 1
seznam rohlik finance 0
lidl seznam seznam 0
idnes novinky - 0
idnes - - 0

22. ALGORITHM
(optimizing
towards goal)
ADSERVER
DSP
…
MASTER DATA
Automated download
YOUR AI
IN CLOUD
PLATFORM
STORAGE
(for free)
USER1; BID 0.2 ;
USER2; BID 0.6 ;
USER3; BID 0.0 ;
USER4; BID 0.9 ;
…
USER1; BID 0.2 ;
USER2; BID 0.6 ;
USER3; BID 0.0 ;
USER4; BID 0.9 ;
…
USER1; BID 0.2 ;
USER2; BID 0.6 ;
USER3; BID 0.0 ;
USER4; BID 0.9 ;
…
HOW OUR AI REALLY WORKS?

23. ALGORITHM
(optimizing
towards goal)
ADSERVER
DSP
…
MASTER DATA
STORAGE
(for free)
USER1; BID 0.2 ;
USER2; BID 0.6 ;
USER3; BID 0.0 ;
USER4; BID 0.9 ;
…
USER1; BID 0.2 ;
USER2; BID 0.6 ;
USER3; BID 0.0 ;
USER4; BID 0.9 ;
…
USER1; BID 0.2 ;
USER2; BID 0.6 ;
USER3; BID 0.0 ;
USER4; BID 0.9 ;
…
TARGETING
PLATFORM
HOW OUR AI REALLY WORKS?
USER / DOMAIN LIST
USER / DOMAIN LIST
USER LIST
AI
CRM
CLOUD
PLATFORM

24. OK, Loan… but what message?
DIGI DATA BASED SALES SIGNALS

25. DIGI DATA BASED SALES SIGNALS

26. DIGI DATA BASED SALES SIGNALS

27. Call script: “Client is interested in taking loan. He
is planning wedding according to his activity on
internet.”
DIGI DATA BASED SALES SIGNALS

28. 50%
Increased effectivity of call center calls in mortgages:

29. • What are the costs?
• What is the operation risk?
• Where are the limits?
• What portfolio can be targeted?
• …

30. • Cheap storage, computational power on demand
• 30% clients training => 100% scoring
• 1st party data (Web paths) => safe
• 3rd party data (Adform paths)=> dying
• …

31. CONTEXT

32. CONTEXT
Case study #3 Similar arcticles/URLs
Keyword based targeting
Topic/Semantic based targeting

33. ENRICH URLS (KEYWORDS, TOPIC, …)
https://www.sport.cz/hokej/extraliga/clanek/1186624-hokej-online-liberec-znovu-unika-sparta-je-na-pokraji-porazky.html

34. We are extending off-line insights by AI
5mil domains...
20 domains
20 keywords
500 domains
500 keywords
All users
Our solution
Regular approach
Uncovering
domains/kws
uknown to
other bidders
word2vec
Vec(king) – Vec(man) + Vec(woman) = Vec(queen)

35. ENRICH URLS
('https://www.sport.cz/hokej/extraliga/clanek/1186624-hokej-online-liberec-znovu-unika-sparta-je-na-pokraji-porazky.html')
https://www.tyden.cz/rubriky/sport/hokej/extraliga/sparta-padla-na-trun-usedl-liberec-zlin-potupil-kladno.html 2.19
https://isport.blesk.cz/clanek/hokej-tipsport-extraliga/373550/sestrihy-hradec-sesadil-spartu-z-trunu.html 1.96
https://www.hokej.cz/dvojnasobny-extraligovy-sampion-se-spartou-znovu-prerusil-karieru 1.61
https://www.ceskenoviny.cz/zpravy/1824498 1.55
OUTPUT

36. Outcome
Regular approach
20 domains
20 keywords
New uncovered
500 domains
500 keywords
Everyone bids for this
(High CPM)
Way less buyers
(Low CPM)

37. Outcome
Regular approach
20 domains
20 keywords
New uncovered
500 domains
500 keywords
Everyone bids for this
(High CPM)
Way less buyers
(Low CPM)
20% lower CPA!

38. ALGORITHM
(optimizing
towards goal)
ADSERVER
DSP
…
MASTER DATA
STORAGE
(for free)
USER1; BID 0.2 ;
USER2; BID 0.6 ;
USER3; BID 0.0 ;
USER4; BID 0.9 ;
…
USER1; BID 0.2 ;
USER2; BID 0.6 ;
USER3; BID 0.0 ;
USER4; BID 0.9 ;
…
USER1; BID 0.2 ;
USER2; BID 0.6 ;
USER3; BID 0.0 ;
USER4; BID 0.9 ;
…
IDNES.CZ; BID 0.4;
ZOZNAM.SK; BID 0.7;
SEZNAM.CZ; BID 0.6;
…
TARGETING
PLATFORM
HOW OUR AI REALLY WORKS?
USER / DOMAIN LIST
USER / DOMAIN LIST
USER / DOMAIN LIST
AI
CRM
CLOUD
PLATFORM

39. • Raw user-level data
• Connection of data sources
• Dynamic environment
Takeaways LIMITS

40. • Raw user-level data
• Connection of data sources
• Dynamic environment
Takeaways LIMITS Takeaways SOLUTIONS
• Transparent vendors
• Building 360
• Platform, intern capability

41. Washingtonova 17/1599
Prague 1, 110 00
Czech Republic
+420 775 556 122
jakub.stech@datasentics.com
Thank you for
your attention
www.datasentics.com
Jakub Stech, Data Science Architect