SAP Fico Training In Hyderabad | SAP Fico Coaching In Hyderabad | SAP Fico Institute In Hyderabad
FICO Machine Learning AML Solution · Responsible for analytic development of FICO’s product and...
Transcript of FICO Machine Learning AML Solution · Responsible for analytic development of FICO’s product and...
![Page 1: FICO Machine Learning AML Solution · Responsible for analytic development of FICO’s product and technology solutions, including Falcon Fraud Manager • 18 years at FICO • Author](https://reader034.fdocuments.in/reader034/viewer/2022042107/5e87499031f5c9751d23468b/html5/thumbnails/1.jpg)
© 2016 Fair Is aac Corporation. Confidential. This pres entation is provided for the recipient only and cannot be reproduced or s hared without Fair Is aac Corporation’s expres s cons ent.
FICO Machine Learning AML SolutionS cott ZoldiChief Analytics Officer, FICO@ S cottZoldi
![Page 2: FICO Machine Learning AML Solution · Responsible for analytic development of FICO’s product and technology solutions, including Falcon Fraud Manager • 18 years at FICO • Author](https://reader034.fdocuments.in/reader034/viewer/2022042107/5e87499031f5c9751d23468b/html5/thumbnails/2.jpg)
© 2016 Fair Isaac Corporation. Confidential. 2
S cott Zoldi, Chief Analytics Officer
• R es pons ible for analytic development of FICO’s product and technology s olutions , including Falcon Fraud Manager
• 18 years at FICO
• Author of 79 patents─ 39 granted and 40 in proces s
• R ecent focus on s elf learning analytics AI for real-time detection of Cyber S ecurity attacks , AML detection, and mobile device analytics
• Ph.D. in theoretical phys ics from Duke Univers ity
![Page 3: FICO Machine Learning AML Solution · Responsible for analytic development of FICO’s product and technology solutions, including Falcon Fraud Manager • 18 years at FICO • Author](https://reader034.fdocuments.in/reader034/viewer/2022042107/5e87499031f5c9751d23468b/html5/thumbnails/3.jpg)
© 2016 Fair Isaac Corporation. Confidential. 3
Money Laundering: The process of creating the appearance that illicit
funds obtained through illegal activity originated from legitimate sources.
![Page 4: FICO Machine Learning AML Solution · Responsible for analytic development of FICO’s product and technology solutions, including Falcon Fraud Manager • 18 years at FICO • Author](https://reader034.fdocuments.in/reader034/viewer/2022042107/5e87499031f5c9751d23468b/html5/thumbnails/4.jpg)
© 2016 Fair Isaac Corporation. Confidential. 4
0.8 – $2T
Source: https://www.unodc.org/unodc/en/money-laundering/globalization.html
![Page 5: FICO Machine Learning AML Solution · Responsible for analytic development of FICO’s product and technology solutions, including Falcon Fraud Manager • 18 years at FICO • Author](https://reader034.fdocuments.in/reader034/viewer/2022042107/5e87499031f5c9751d23468b/html5/thumbnails/5.jpg)
© 2016 Fair Isaac Corporation. Confidential. 5
$16M2004
$8.9B2014556 X
![Page 6: FICO Machine Learning AML Solution · Responsible for analytic development of FICO’s product and technology solutions, including Falcon Fraud Manager • 18 years at FICO • Author](https://reader034.fdocuments.in/reader034/viewer/2022042107/5e87499031f5c9751d23468b/html5/thumbnails/6.jpg)
© 2016 Fair Isaac Corporation. Confidential. 6
human trafficking,the second most profitable
crime after narcotics
terroristfinancing
narcotics trafficking
![Page 7: FICO Machine Learning AML Solution · Responsible for analytic development of FICO’s product and technology solutions, including Falcon Fraud Manager • 18 years at FICO • Author](https://reader034.fdocuments.in/reader034/viewer/2022042107/5e87499031f5c9751d23468b/html5/thumbnails/7.jpg)
© 2016 Fair Isaac Corporation. Confidential. 7
F ICO® Falcon® Fraud Manager
Falconintroduced
20
15
10
5
01990 1994 1998 2002 2006 2014
Fraud Losses
2010
Percentage of the US payment cards protected by FICO fraud solutions90%Banks participating in FICO’s fraud data cons ortium – driving ins ight and analytic innovation
9,000
Active financial accounts protected by FICO worldwide – providing a global pers pective2.6B
Average res pons e time for fraud decis ions rendered by FICO’s low-latency real-time engine
10ms
![Page 8: FICO Machine Learning AML Solution · Responsible for analytic development of FICO’s product and technology solutions, including Falcon Fraud Manager • 18 years at FICO • Author](https://reader034.fdocuments.in/reader034/viewer/2022042107/5e87499031f5c9751d23468b/html5/thumbnails/8.jpg)
© 2016 Fair Isaac Corporation. Confidential. 8
Combating Money Laundering Today
o As certain compliance ris ko KYCo Obs erve-and-R eporto S AR so S ubjectiveo R ule-bas ed
![Page 9: FICO Machine Learning AML Solution · Responsible for analytic development of FICO’s product and technology solutions, including Falcon Fraud Manager • 18 years at FICO • Author](https://reader034.fdocuments.in/reader034/viewer/2022042107/5e87499031f5c9751d23468b/html5/thumbnails/9.jpg)
© 2016 Fair Isaac Corporation. Confidential. 9
“Increas ingly, regulators recognize that rules alone are not an effective manner of detection and are pres s uring banks to include more s ophis ticated analytics .”
Aite Group LLC, “Global AML Vendor Evaluation” 2015
![Page 10: FICO Machine Learning AML Solution · Responsible for analytic development of FICO’s product and technology solutions, including Falcon Fraud Manager • 18 years at FICO • Author](https://reader034.fdocuments.in/reader034/viewer/2022042107/5e87499031f5c9751d23468b/html5/thumbnails/10.jpg)
© 2016 Fair Isaac Corporation. Confidential. 10
Vocabularies to des cribe trans action behavior
Think of trans action behavior and events as words from a vocabulary
The s tream of behavior is s een as the s equence of words
Current Account
AmountsWire Trans fer Country
Acces s Channel
Example word:
“$500-$750_Wire_FR A”
Patents 14/796,547 (USA) ,14/613,300 (USA)
Word“$500-$750_Wire_FRA” “$100-250_EFT”
“$250-$500_EFT”
“$500-$750_Wire_ITA”
![Page 11: FICO Machine Learning AML Solution · Responsible for analytic development of FICO’s product and technology solutions, including Falcon Fraud Manager • 18 years at FICO • Author](https://reader034.fdocuments.in/reader034/viewer/2022042107/5e87499031f5c9751d23468b/html5/thumbnails/11.jpg)
© 2016 Fair Isaac Corporation. Confidential. 11
Learning archetypes from trans actions : Collaborative Profiling
• Bayesian Learning─ Uns upervis ed─ Learn archetypes from
millions of cus tomers .
Patents 14/796,547 (USA) ,14/613,300 (USA)
Cus tomer’s data s tream:
From many other cus tomers
Learned Archetypes (~10’s )
![Page 12: FICO Machine Learning AML Solution · Responsible for analytic development of FICO’s product and technology solutions, including Falcon Fraud Manager • 18 years at FICO • Author](https://reader034.fdocuments.in/reader034/viewer/2022042107/5e87499031f5c9751d23468b/html5/thumbnails/12.jpg)
© 2016 Fair Isaac Corporation. Confidential. 12
Clus tering archetypes : Mis alignment with clus ters is s us picious
Archetype 1
Archetype 2
Archetype n Scenario: Existing customer moves out of cluster:• Sleeper account
being activated
Patents 15/074,856 (USA) ,14/074,977 (USA)
Score 120
Score 640
340
520
![Page 13: FICO Machine Learning AML Solution · Responsible for analytic development of FICO’s product and technology solutions, including Falcon Fraud Manager • 18 years at FICO • Author](https://reader034.fdocuments.in/reader034/viewer/2022042107/5e87499031f5c9751d23468b/html5/thumbnails/13.jpg)
© 2016 Fair Isaac Corporation. Confidential. 13
R eal-World AML Example: S AR dis tribution in archetype s pace
• Many S AR s are outliers from normal cus tomers along certain archetypes
Arc
hety
pe 1
2
Archetype 10
S AR Cus tomer
Normal cus tomer
Archetype Dis tribution
![Page 14: FICO Machine Learning AML Solution · Responsible for analytic development of FICO’s product and technology solutions, including Falcon Fraud Manager • 18 years at FICO • Author](https://reader034.fdocuments.in/reader034/viewer/2022042107/5e87499031f5c9751d23468b/html5/thumbnails/14.jpg)
© 2016 Fair Isaac Corporation. Confidential. 14
Autoencoders for uns upervis ed anomaly s coring
• Autoencoders are deep neural nets trained to repres ent/compres s input by minimizing recons truction error.
Input layer
Hidden layer
Output Layer
Target: Input
Train to minimize difference between input and output through “bottleneck layers ”
![Page 15: FICO Machine Learning AML Solution · Responsible for analytic development of FICO’s product and technology solutions, including Falcon Fraud Manager • 18 years at FICO • Author](https://reader034.fdocuments.in/reader034/viewer/2022042107/5e87499031f5c9751d23468b/html5/thumbnails/15.jpg)
© 2016 Fair Isaac Corporation. Confidential. 15
Autoencoders : R econs truction error meas ures s imilarity to training data
Input layer
Hidden layer
Target: InputOutput Layer
R econs truction error
Patent PCT/US15/63395
• For anomaly scoring, this reconstruction error indicates how much a sample differs from the training population.
![Page 16: FICO Machine Learning AML Solution · Responsible for analytic development of FICO’s product and technology solutions, including Falcon Fraud Manager • 18 years at FICO • Author](https://reader034.fdocuments.in/reader034/viewer/2022042107/5e87499031f5c9751d23468b/html5/thumbnails/16.jpg)
© 2016 Fair Isaac Corporation. Confidential. 16
R eal-world AML application: Autoencoder finds outlier in archetype s pace
• Autoencoder trained on Collaborative Profiling archetypes
• High s cores when autoencoder finds archetype mixtures very different from training s et.
Autoencoder
![Page 17: FICO Machine Learning AML Solution · Responsible for analytic development of FICO’s product and technology solutions, including Falcon Fraud Manager • 18 years at FICO • Author](https://reader034.fdocuments.in/reader034/viewer/2022042107/5e87499031f5c9751d23468b/html5/thumbnails/17.jpg)
© 2016 Fair Is aac Corporation. Confidential. This pres entation is provided for the recipient only and cannot be reproduced or s hared without Fair Is aac Corporation’s expres s cons ent.
Thank you!