Teradata Partners Conference 2017

1. © 2015 Teradata
Dr Patrick Deglon
Vice President, Advanced Analytics, Teradata
A Journey into bringing (Artificial) Intelligence
to the Enterprise

2. 2
After a PhD in Particle Physics and ten years at the University of
Geneva studying the creation of the Universe, Patrick spent the next
decades driving business insights at eBay, Motorola Mobility, and Teradata.
At eBay, he led significant improvements in marketing effectiveness by developing
methods to measure incremental sales, and by running large scale experiments on
Internet marketing channels.
At Motorola Mobility, he raised the bar in Analytics and on-boarded open Google
tools and technologies.
In Dec 2016, he joined Teradata as the Vice President of Advanced Analytics driving
the strategy, direction, investment and realization of Teradata’s advanced analytics
portfolio, including the Teradata Database, Aster Analytics, and Open Source
Software.
He is married with two kids and recently moved to San Diego, California, USA.
Dr Patrick Deglon Bio

3. 3
Image: CERN
Once upon a time…

4. 4
4
Geneva
Switzerland
Image: CERN
27 km underground tunnel
for the LEP & LHC accelerator
Mont
Blanc

5. 5
Images: CERN

6. © 2017Image: CERN
Image: Weizmann Institute

7. 7
Tape robot
Source: CERN
PAW – Physics Analysis Workstation
Source: Wikipedia
Data collection & analysis was done in
Fortran. Advance analysis/statistics was
done through PAW. [1996-2002]

8. 8
Example of particles collision

9. 9
Solving the puzzle… which particles go together?
?
A
B
CD
1. AB + CD?
2. AC + BD?
3. AD + BC?

10. 10
Solution: Big Data infrastructure enables large scale
computational such as combine all possibilities (cross-product)
Statistical Noise
Signal
(particle resonance)
Source: http://www.atlas.ch/news/2011/ATLAS-discovers-its-first-new-particle.html
Schematic View CERN Example
(discovery of a new particle bb)

11. 11
Where are we going?

12. 12
Evolution of mankind
Homo
Sapiens
Homo
erectus
Homo
habilis
Human-like
apes

13. 13
Evolution of mankind
Homo
Sapiens
Homo
Technicus
Homo
erectus
Homo
habilis
Human-like
apes

14. 14
Evolution of mankind
1973: First hand-held portable telephone
1989: Web proposal
2009: First microchip implant
...
Homo
Sapiens
Homo
Technicus
Homo
erectus
Homo
habilis
Human-like
apes

15. 15
The world is changing…
It’s estimated that between
35%and 50%of jobs
that exist today are at risk of
being lost to automation.
Forbes – Mar 3rd 2017 (link)
The Associated Press expanded its quarterly
earnings reporting from approximately 300
stories to 4,400with the help of AI-
powered software robots.
HBR – Nov 2nd 2016 (link)
78% of the surveyed managers believe
that they will trust the advice of intelligent
systems in making business decisions in the
future.
HBR – Nov 2nd 2016 (link)
Digital is the main reason just over 50%
of the companies on the Fortune 500 have
disappeared since year 2000.
Pierre Nanterme, CEO Accenture
JPMorgan Software Does in
Seconds What Took Lawyers
360,000 Hours
Bloomber – Feb 27th 2017
Google Researchers Are Teaching Their AI
to Build Its Own, More Powerful AI
Science Alert, May 2017
75% of S&P 500 incumbents
will be gone by 2027.
McKinsey – July 2014 (link)

16. 16
Navigating the next industrial revolution
1784 1870 1969 Now
1st Revolution
Mechanization, Water
Power,
Steam Power
2nd Revolution
Mass production,
Assembly line,
Electricity
3rd Revolution
Computer
and Automation
4th Revolution
Cyber Physical
Systems

17. 17
Self-driving cars/trucks (e.g. Volvo)
Customer Care bots (e.g. Barclays)
Cashier-free stores (e.g. Tesco)
Personal Assistant (e.g. Apple)
Life Science (e.g. Kaiser Permanente)
Automated Incident Diagnose
& Recovery (e.g. Siemens)
Machine
Learning
Business
Rules
+
Cyber Physical Systems: The Rise of the Machines

18. 18
• Product Development
• Manufacturing
• Operations
• Information Technology
• Customer Supports
• Human Resources
• Legal
• Sales & Marketing
• Finance & Strategy
Data Insights
Analytics
Execution
Decision
Making
The Role of Cyber Physical
Systems in the Enterprise

19. 19
Prescriptive
Analytics
Predictive
Analytics
Descriptive
Analytics

20. 20
Analytics Evolution
Descriptive
Predictive
Prescriptive
What is happening?REPORTING
QUERY & DRILL DOWN
ALERTS
MACHINE LEARNING
DEEP LEARNING
What exactly is the problem?
What actions are needed?
Self-learning systems with
regression.
Deep neural networks.
FORECASTING What are the hidden patterns?
What will happen next?
SIMULATION What could happen?
2020s
2000s
1980s
Copyright 2017 Teradata

21. 21
2000s: The glory days
of digitalization
2

22. 22
Case study: Online Search
• #TeradataUniverse #AIEnterprise

23. 23
Case study: Online Search
• #TeradataUniverse #AIEnterprise

24. 24
Case study: Online Search
• #TeradataUniverse #AIEnterprise
Natural/Organic
Search (free)
Paid Search

25. 25
Jan
1st
Feb
1st
Mar
1st
Customer behaviors and Internet Marketing Investment
25

26. 26
Jan
1st
Feb
1st
$ $ $ $ $ $ $
Mar
1st
Customer behaviors and Internet Marketing Investment
26
Behavioral purchases
Uncorrelated to Marketing

27. 27
Jan
1st
Feb
1st
$ $ $ $ $ $ $
click
$ $ $
Behavioral purchases
Uncorrelated to Marketing
click Mar
1st
$
Influenced purchase
Correlated to Marketing
Customer behaviors and Internet Marketing Investment
27

28. 28
Jan
1st
Feb
1st
$ $ $ $ $ $ $
click
$ $ $
Behavioral purchases
Uncorrelated to Marketing
click Mar
1st
$
Influenced purchase
Correlated to Marketing
Which customer purchases are
influenced by Marketing?
Customer behaviors and Internet Marketing Investment
28

29. 29
X days

2 purchases
missing
X days

Y days
all purchases
are incremental1 purchase is
uncorrelated
Y days

Jan
1st
Feb
1st
$ $ $ $ $ $ $
click
$ $ $
Behavioral purchases
Uncorrelated to Marketing
click Mar
1st
$
Influenced purchase
Correlated to Marketing
Which customer purchases are
influenced by Marketing?
Customer behaviors and Internet Marketing Investment
29

30. 30
Let’s us a standard method from Particle Physics
?
A
B
CD
1. AB + CD?
2. AC + BD?
3. AD + BC?

31. 31
Solution: Big Data infrastructure enables large scale computational
such as combine all possibilities (cross-product)
Statistical Noise
Signal
(particle resonance)
Schematic View

32. 32
Solution: Big Data infrastructure enables large scale computational
such as combine all possibilities (cross-product)
Statistical Noise
Signal
(particle resonance)
Schematic View CERN Example
(discovery of a new particle bb)

33. 33
Solution: Big Data infrastructure enables large scale computational
such as combine all possibilities (cross-product)
Statistical Noise
Signal
(particle resonance)
Schematic View
Combine correlated events and uncorrelated events produce a system with a
statistical noise (which is simple enough to extract) and the researched signal
CERN Example
(discovery of a new particle bb)

34. 34
Positive Latency
Purchase after Click (potential causality)
Behavior & Internet Marketing impact
Latency (days)
0 2 4 6 8 10 12 14
User clicks on an
ad-banner at time=0
User makes a purchase
7 days later
Latency time for each pair click - purchase

35. 35
Positive Latency
Purchase after Click (potential causality)
Behavior & Internet Marketing impact
Latency (days)
Number of events
(pairs click-
purchase)
Negative Latency
Purchase before Click (no causality)
Behavior only
-14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10 12 14
User clicks on an
ad-banner at time=0
User makes a purchase
7 days later
Latency time for each pair click - purchase

36. 36
Positive Latency
Purchase after Click (potential causality)
Behavior & Internet Marketing impact
Level of
behavioral
purchases
Latency (days)
Number of events
(pairs click-
purchase)
Negative Latency
Purchase before Click (no causality)
Behavior only
-14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10 12 14
User clicks on an
ad-banner at time=0
User makes a purchase
7 days later
Latency time for each pair click - purchase

37. 37
Level of
behavioral
purchases
Positive Latency
Purchase after Click (potential causality)
Behavior & Internet Marketing impact
Level of
behavioral
purchases
Latency (days)
Number of events
(pairs click-
purchase)
Negative Latency
Purchase before Click (no causality)
Behavior only
-14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10 12 14
User clicks on an
ad-banner at time=0
User makes a purchase
7 days later
Latency time for each pair click - purchase

38. 38
Marketing
incrementality
(correlated
purchases) Level of
behavioral
purchases
Positive Latency
Purchase after Click (potential causality)
Behavior & Internet Marketing impact
Level of
behavioral
purchases
Latency (days)
Number of events
(pairs click-
purchase)
Negative Latency
Purchase before Click (no causality)
Behavior only
-14 -12 -10 -8 -6 -4 -2 0 2 4 6 8 10 12 14
User clicks on an
ad-banner at time=0
User makes a purchase
7 days later
Latency time for each pair click - purchase

39. 39
Don’t
Do Marketing
Do Marketing
Marketing 101

40. 40
Don’t
Do Marketing
Do Marketing
No Purchase
Purchase
Marketing 101

41. 41
Don’t
Do Marketing
Do Marketing
No Purchase
Purchase L L
Marketing 101

42. 42
Don’t
Do Marketing
Do Marketing
No Purchase
Purchase L L
D D
Marketing 101

43. 43
Don’t
Do Marketing
Do Marketing
No Purchase
Purchase L L
D D
C
C
Marketing 101

44. 44
Don’t
Do Marketing
Do Marketing
No Purchase
Purchase L L
D D
C
C
?
?
Marketing 101

45. 45
Don’t
Do Marketing
Do Marketing
No Purchase
Purchase L L
D D
C
C
?
?
Cost
Direct Return
Incr Return
Marketing 101

46. 46
Don’t
Do Marketing
Do Marketing
No Purchase
Purchase L L
D D
C
C
?
?
Cost
Direct Return
Incr Return
Rule #1: Never, ever, spend money
unless you really-really have to
Marketing 101

47. 47
Output
Return (Revenues)
Marketing 101

48. 48
Output Cost
Return (Revenues)
Marketing 101

49. 49
Output Cost
Return (Revenues)
Marketing 101
Profit

50. 50
Output Cost
Return (Revenues)
Marketing 101
Max Sales,
No Profit
Total
ROI = 0
Profit

51. 51
Output Cost
Return (Revenues)
DReturn = DInvestment
i.e. marginal ROI = 0
Marketing 101
Max Sales,
No Profit
Total
ROI = 0
Max Profit Profit

52. 52
Output Cost
Return (Revenues)
DReturn = DInvestment
i.e. marginal ROI = 0
Marketing 101
Max Sales,
No Profit
Total
ROI = 0
Max Profit
Rule #2: If you have to spend, you spend
to the point of marginal return=0
Profit

53. 53
• How did all this started?
2020s: The rise
of the machines

54. 54
Teradata market research
90% of customers want to analyze
their data in-database
85% of advanced analytic workloads are run outside
of Teradata
50% of the data scientists primarily do their data preparation
on a laptop, desktop, or company server rather than using in-database
functions
70% of Data Science project time is spent in
finding and preparing the data

55. 55
There is an increasing proliferation of analytic tools
Analytic Engines Data FormatsLanguagesWorkbenches
voice

56. 56
Teradata
Analytics Platform
with In-Database Analytic Functions
and Analytic Engines
Flexible User Interfaces,
Languages and
Workbenches
Analytical
Ecosystem
AnalyticalCapabilities
Data Formats, Types and Sources
From “Data Warehouse” to Leading “Analytics Platform”
Non-relational
Data Types and
File Systems
* Anticipated future capabilities
* *

57. 57
Analyze Anything!
* Anticipated future capabilities
Preferred Tools
and Languages
Support for
Multiple Data Types
The Best Analytic Functions
and Engines
T E R A D A T A A N A L Y T I C S P L A T F O R M
Data Science Workbenches – Jupyter,
RStudio, KNIME, SAS, Dataiku, and Teradata
Profiler*
Languages - SQL, SAS, Python, R, C, Java,
C++, JavaScript*, Go*, Perl*, and Swift*
Model Management: AnalyticOps*
Data formats - JSON, BSON*, AVRO*, CSV,
XML, PDF*, Voice*, Video*, and Images*
Data types - Interval, Geospatial, Temporal,
and Time Series
Data sources - Integrate with Hadoop* and
S3*
Teradata Functions - Time Series,
Attribution, Statistics, Data Transformation,
Path & Pattern Analysis, Association,
Clustering, Decision Trees, Naïve Bayes, Deep
Learning, Text Mining, Graph, and Location
Analysis
Analytic Engines – SQL, Graph, Machine
Learning, R, Python, Spark* and Deep
Learning* (e.g. TensorFlow, Gluon, Theano)

58. 58
Teradata Analytics Platform
Connectors & APIs
Teradata
Engine
Aster
Engines
Spark
Engine*
Deep Learning
Engine*
Custom
Engine*
Self-Service Tools,
Languages & Apps
Analytic Functions, Engines and Data Types
* Anticipated future capabilities

59. 59
Teradata Analytics Platform
Teradata
Engine
Persistent
Storage
SQL Access
TD Studio, BI and
Visualization tools
Languages
SQL, SAS, Python, R,
C, Java
Tools
Jupyter, RStudio,
SAS
Hadoop
S3
Others
Aster
Engines

60. 60
Teradata Analytics Platform
Teradata
Engine
Persistent
Storage
Cross-Engine
Orchestration
SQL Access
TD Studio, BI and
Visualization tools
Hadoop
S3
Others
Languages
SQL, SAS, Python, R, C,
Java, C++, C#,
JavaScript
Tools
Jupyter, RStudio,
KNIME, SAS, Dataiku,
Profiler
* Anticipated future capabilities
Aster
Engines
Spark
Engine*
Deep
Learning
Engine*
Custom
Engine*

61. 61
User story example: Utility company wants to build a better predictive
model of transformer failures
COO,
Kathy
Data Scientist,
Jason
Jason, we lost $50M last quarter due to
transformer failures that were not
predicted.
Can you build a better predictive
model to forecast transformer failure?

62. 62
Better predictive model of transformer failure is easy
to create with TD advanced analytics
AI Incubator
Mid 2018
Future (TBC)
Feature availability:
• Jason access his favorite tool Jupyter hosted in
AppCenter, collocated with the data
• He explores the data using in-database functions
exposed through the Teradata Python package
and visualization in Jupyter
• Running in-database descriptive functions,
he select the 80 most interesting
features.
• Jason ingests external log files and weather
data located in the company data lake
• He use the Knowledge Sharing App to
understand the logic of the previous model
• Jason uses in-database function to clean and
append weather & log data, and flag failures
• He creates an analytic table in the database
without having to move the data
• Jason writes in Python a simple TensorFlow
code that he trains on his Jupyter session in the
Analytics platform collocated with his dataset
• He further extend the complexity of the model
and train & test his model on the full dataset in the
distributed and parallel TensorFlow Engines
• Jason deploys the model in the
AnalyticOps that creates a new version of
the predictive failure model family.
• The AnalyticOps promotes the model to
production, as it’s 3x more accurate
• Jason works with IT to build a web app on
AppCenter to help business identify
bad transformers that are likely to fail
AnalyticOps
AppCenter
QueryGrid
Aster functions
Data Lab
Python & R
packages
Product.
Data
Prep
Explora-
tion
AppCenter
Training

63. 63
Teradata Analytics Platform Key Customer Value Prop
• It’s not a new product – it is the future of our core technology: Integrated in Database SW
license 16.20+
• Uniqueness of Teradata Database: resiliency, scalability, security, user management, …
• Depth of Aster Analytics functions: Statistics, Data Transformation, Path/Pattern Analysis,
Association, Clustering, Decision Tree, Naïve Bayes, Neural Networks, Text Mining, Graph
Analysis, Location Analysis
• New multi-genre Analytics engines: Aster Graph engine, Aster Map/Reduce engine, Aster R
Engine, technologic platform for future engines (Spark, Deep Learning, …)
• Wealth of libraries/plug-ins for end-user tools and languages: SQL, RStudio, Jupyter, Python,
KNIME, Dataiku, SAS, Teradata AppCenter, BI tools, Teradata Studio, and others

64. Thank You!
Rate This Session #
with the PARTNERS Mobile App
0189
Follow Me
Twitter @pdeglon
LinkedIn www.linkedin.com/in/deglon
Questions/Comments
Email: Patrick.Deglon@Teradata.com