How can we take the state of the art in deep learning and AI research, and transplant it into a large bank to deliver useful results which impact the general public? To answer this broad-reaching question, we take the viewer through a solution Think Big Analytics recently deployed at a major European bank for fraud detection, using state of the art AI techniques and a near-real time open-source architecture. We show how financial transactions can be transposed into a form where the latest AI techniques in image recognition can be leveraged, in surprisingly novel ways. We have been able to more accurately detect fraud and reduce financial crime, cutting losses and improving customer experience. We describe some architectures which can be used to do this in production, at scale, in global financial institutions. Speaker: Tim Seears, Director of Data Science, Think Big Analytics, a Teradata Company

1. Fighting Financial Crime
with Artificial Intelligence
Tim Seears
Area Practice Director, Data Science – Asia Pacific & Korea

2. 2
So far we have delivered
150+ successful projects for
100+ clients worldwide.
500+ employees
(2017)
Vendor-neutral with an
open source focus.
Full spectrum consulting, data
engineering, data science &
support.
Apache
Hadoop
and cloud
ecosystem
integration.
Founded in 2010
industry thought
leader.
Fixed fee offerings
for data science and
engineering.
Who Is Think Big?
1st Big Data provider 100%
focused around open source.

3. 3
Help customers be financially confident
and achieve their ambitions by making
daily banking and important financial
decisions easy.
19,000
employees
(2017)
1,800 corporate and
institutional customers
Covering Personal Banking, Business
Banking, Corporates & Institutions and
Wealth Management.
2.7 million
personal
customers
236,000 small and
medium-sized
business customers
Making banking
easier for over 145
years.
Who is the customer?
Leading Nordic universal bank with strong
local roots and bridges to the rest of the
world

4. 4
PayPal WalmartJohn Deere Lowe’s Wells FargoJP Morgan
This is Happening in the Enterprise today

5. 5
“
’’
Over the next decade, AI won’t replace
managers, but managers who use AI
will replace those who don’t.
HBR July 2017

6. 7
Data Driven Approach to Fight Fraud
Fast evolving fraud sophistication, AI
Ambitions for Fraud Project
Become client’s advanced
analytics blueprint
Data driven approach to real
time scoring of transactions
Reduce false-positives &
Enhance fraud detection rate
ONLY ~40%
of fraud cases are detected
Low Detection Rate
99.5%
of cases are not
fraud related
Many false positives
Challenges for Fraud Detection
Tens of Millions
€ lost each month
High Fraud Loss
© 2017 Teradata

7. 8
Fraud Types – Customer Initiated
Nigeria/
Investment Scam
CEO
Fraud
Customer
Initiated
© 2017 Teradata
Beneficiary Account
Change
Fake
Invoice
Rental Scam/
Goods Not
Received

8. 9
ID Theft
SPEAR
Phishing
Vishing/Support Scam
Malware
Phishing/Smishing
Fraud Types – Fraudster Initiated
Fraudster
Initiated
© 2017 Teradata

9. 10
Modeling
Challenges
© 2017 Teradata
• Class imbalance
(100,000:1 non-fraud vs. fraud)
• Assigning fraud labels from
historic data
• Fraud is ambiguous
• Not all features available in
real-time
• Most machine learning sees
transactions atomically

10. 11
Advanced Platform for Fraud: Data-Driven Approach
How It Works:
• Understanding the domain
• Gathering and preparing the data
• Automatically generating the rules and recognizing fraudulent patterns by
training models on historical data
• Automatically maintaining the engine by retraining the model
Pros:
• Automatic/data-driven/objective
inference of the rules
• Ability to detect patterns in a high
dimensional data input
• Fast detection of new/changing
fraudulent patterns
Cons:
• Might be unintuitive and hardly
interpretable
• Data preparation and feature aggregation
is time consuming
© 2017 Teradata

11. 12
Phase One
© 2017 Teradata

12. 13
Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul
Project kick-off
30th September
Teamwork to Deliver Value in Each Project Iteration
2016 2017
Kick-off: Data
Scientist track
Go-Live plan set
for the project
19th December
Kick-off:
Engineering track
Models successfully in
Shadow Production
4th March
First round trip
test transaction
First production
virtual machine
Production Hardening
for HA and Security
Full productionisation of
the Fraud Engine
Cross-
functional
Collaboration
Cont. Deep
Learning
modeling
© 2017 Teradata

13. 14
>
Banking
Anti-Fraud
Solution
© 2017 Teradata
Banking Anti-Fraud Solution
By leveraging the power of a
thoughtful and strong data
and analytics strategy, we
unleash high impact business
outcomes.
• Multiple models running in production
at the same time
• Mix of traditional and advance deep
learning methods
• AnalyticsOps: Deploying machine
learning models in production
Model Management
Framework
• Organisation and silos of data
• Real-time data integration
• Security and Procedures: following
existing bank procedures
Data Modelling, Pipeline
and Ingestion
• Hard to operationalize insights
• Availability of analytic capabilities/ skills
and data
• Interpreting the results of machine
learning models
Machine Learning and
Artificial Intelligence

14. 15
Advanced Platform for Fraud: Data-Driven Approach
Manual
Evaluation
eBanking
Mobile
Pay
Business
Online
Global Payment
Interface
Fraud?
Create
Verification
Fraud
Payment
Weblog Data
Basic Customer
Data
Historic
Transactions
Customer Product
Data
Aggregated
Customer Data
Fraud?
Advanced Analytics
Platform for Fraud
Central Fraud Engine
Process Payment to
Beneficiary
Yes/Maybe
No
Yes !
Return Payment
to Customer
Update
Fraud Data
© 2017 Teradata
No

15. 16
Model Management
Real Time Scoring
System
Production
Operations
Source
Systems
Data
Managers
Framework to Enhance AnalyticsOps Capabilities
Log
Transactions
Add Historic
Data
Model
Performance
Promote
Models
© 2017 Teradata
Data Scientists
Development & Deployment
Management
Data Science Lab
Data Analysis &
Model Development
Data
Management
Developers &
Release Managers

16. 17
Key Requirement: Model Interpretation
• We have deployed LIME (Locally Interpretable Model Explanation) for
customers
– Improves trust
– Compliance with EU’s General Data Protection Regulation (GDPR)
© 2017 Teradata
17.6% Fraud Probability
Customer Amount
Debit Amount
Avg. # Trans Cred Acc
# Prev to This Dest
# Xfer Accounts
X% score due to:
+ transfer amount
+ destination country
+ last year monthly spend
What features are most important to
this decision?

17. 18
Machine Learning Results
(Live System: 60 transactions/sec.)
Ensemble of boosted decision trees and
logistic regression.
From online validation of the model:
● 25-30% false positive reduction, with
over 35% increase in detection rate
● Opportunity to expand model with
additional features, retrain on recent data
and add additional models to
the ensemble.
● Models can be expanded to additional
channels
Rule Engine on
validation set
© 2017 Teradata

18. 19
Deep Learning
© 2017 Teradata

19. 20
Current models can
only catch ~70% of all
fraud cases
Deep Learning Opportunity
Traditional ML models view
transactions atomically
Often missed fraud
transactions are part of
a series
Capturing correlation
across many features
© 2017 Teradata

20. 21
Three Deep Learning Architectures to Deliver Value
• Designed for spatial
correlated features, but
by transforming
transactions into a
2D image, we can
learn temporal
correlated features.
• Deeper ConvNet allows
learning more complex
& general features.
Goal: Learn kernels from
temporal & static features
to gain insight into the
characteristics of fraud.
• Learn temporal
information and classify
if the sequence of
transactions
contains fraud.
• Shares knowledge
across learning time.
Goal: Learn transaction
patterns within a window.
Two solutions can be
tested: flag fraud or predict
next transaction and define
an error.
• Learn how to generate
normal transactions,
potentially large volumes
of non-fraud data.
• AE provide a low level
representation of the
data.
Goal: Build a model that
learns how to generate
non-fraud data. To detect
fraud, define a
reconstruction error rate for
the fraud cases
Auto-Encoders
LSTM
ConvNet
© 2017 Teradata

21. 22
How Can We Create an Image From Bank Transactions?
t0 X_0, X_1, ... X_n
dt
t1 X_0, X_1, ... X_n
t2 X_0, X_1, ... X_n
ts X_0, X_1, ... X_n
...
Top k Features Correlation
...
X_0
X_41 X_5
X_30
X_29X_31X_10
X_37
X_3
X_1
X_42 X_40
X_32
X_15X_35X_2
X_16
X_31
X_2
X_3 X_15
X_4
X_1X_28X_40
X_31
X_49
X_n
X_26 X_9
X_40
X_35X_28X_2
X_17
X_1
...
X_0
X_41 X_5
X_30
X_29X_31X_10
X_37
X_3
X_1
X_42 X_40
X_32
X_15X_35X_2
X_16
X_31
X_2
X_3 X_15
X_4
X_1X_28X_40
X_31
X_49
X_n
X_26 X_9
X_40
X_35X_28X_2
X_17
X_1
...
...
...
...
...
X_0
X_41 X_5
X_30
X_29X_31X_10
X_37
X_3
X_1
X_42 X_40
X_32
X_15X_35X_2
X_16
X_31
X_2
X_3 X_15
X_4
X_1X_28X_40
X_31
X_49
X_n
X_26 X_9
X_40
X_35X_28X_2
X_17
X_1
N
dt
t0
t1
ts
Input Output
Raw Features
Add correlated features in a
clock-wise manner
© 2017 Teradata
Image size is:
[10 x 3, 50 x 3, 1]

22. 23
Convolutional Layers for Trans2D
Kernels
of 3x3
...X_0 X_1 X_2 X_n
...X_0 X_1 X_2 X_n...
...
...
...
...X_0 X_1 X_2 X_n
Features
Time
Strides of 3
Strides
of 1
First Convolutional Layer Architecture
© 2017 Teradata

23. 24
2D Transaction Image Example
© 2017 Teradata
Non-fraud Transaction Image
Non-fraud
Fraud Transaction Image
X-axis: features, Y-axis: time
Fraud
Non-fraud

24. 25
Network Architecture for CNNs
Fraud
Normal
50
30
25
15
13
8CNN
Fraud
Normal50
30
25
15
25
15
25
15
25
15
© 2017 Teradata

25. 26
Inside the ResNet model
64 Filters
Activations After
the CNN
Residual Blocks
FraudNon-fraud
© 2017 Teradata

26. 27
Deep Learning
First Results
on the fraud verification dataset
Comparison of the three deep learning
models andthe traditional machine
learning ensemble model.
© 2017 Teradata
• Ensemble model (AUC 0.89)
• ConvNets (AUC 0.95)
• LSTM (AUC 0.90)
• ResNet (AUC 0.94)

27. 28
Lessons Learned: Take-Aways From client project
Deep
learning
adoption
from pictures
to financial
transactions
Enhancement of
data quality &
cluster
capabilities with
data ingestion
Building
Analytics
Ops
capabilities
to support
business
units
Leveraging
experience
from Fraud
advanced
analytics to
deliver extra
use cases
© 2017 Teradata

28. 29
Success: From PowerPoint to production in 8 sprints
Team effort: Thorough collaboration across IMD, GFU and Think Big
Synergy: Successfully spearheaded innovation in all involved systems
Inspiration: Bank advanced analytics blueprint sets a generic scene for combatting new
challenges in advanced analytics
Agile influence: Using an agile approach we were able to quickly deliver within the
challenging timeframe.
© 2017 Teradata
Big Data Team – Lessons Learned

29. 3030 © 2017 Teradata

30. 31
Appendix
© 2017 Teradata

31. 32
Advanced Analytics Fraud Platform
• Develop a scalable and expandable platform which follows the Bank’s blueprint of
digitalization
• 100 % data-driven approach to find patterns in the data and complement the
existing fraud engine
• Use Hadoop to handle the large data volumes for training models on transaction
data
• Implement a real-time solution that can score live transactions such as
e-banking, credit card and mobile payments.
• Reduce amount of false-positives by at least 20-40 %
© 2017 Teradata

32. 33
Ambitions for the Advance Analytics project
• The goal is to enhance fraud detection and
reduce false positives across products.
• Fraud is only the first of many possible
advanced analytics use cases.
• The project leads to the creation of the Bank’s
advanced analytics blueprint.
• Bank’s ambition is to become one of the bank
leading in advanced analytics capabilities.
Challenges in Fraud Detection at client
• Low detection rate: Existing human-written
rule engine catches ~40% of fraud cases.
• Many false positives: 99.5% of all cases
investigated are not fraud related.
• High fraud loss: Tens of millions of € total
fraud per month.
• Mobile payments are quickly growing in
number.
• Fraud evolving rapidly, with increased
sophistication: Bank must modernize its anti-
fraud arsenal.
© 2017 Teradata