Downloaded 21 times
![This content is classified as Internal
18
send
money
Level 02 – Graph rule based approach
0.8 0.3
Org.
0.4
user 01 user 02
user 05
user 03
send
money
. . .
user N
send
money
0.93
0.2
user 06
0.3
• Network anomalies are hard to model and to check
with relational database
• Graph databases, like Neo4j, are the best choice when
your analytical focus is link between peers, that is a
network.
• Thanks to graph data structure several anomaly
detection rules can be built to fight groups of
suspicious transactions or users.
MATCH
(u:User)-[:send_money*]->(v:User)
WHERE
u.score > 0.8 and v.score > 0.8
Level 02 Description](https://image.slidesharecdn.com/20230207datascienemilannexi-230216113550-2bd6a8b0/75/ML-Graph-algorithms-to-prevent-financial-crime-in-digital-payments-18-2048.jpg)
![This content is classified as Internal
19
Level 02 – Graph rule based approach
Data
Lake
ingest
AML
Feature
Store
Amazon
S3
Amazon
Glue
EC2
load
Graph
anomalies
Back-end
has
account
Front-end
query
process
DATA MODEL
1
Neo4j CE on EC2 (not managed)
2
Neo4J Spark connector
3
MATCH
(u:User)-[:send_money*]->(v:User)
WHERE
u.score > 0.8 and v.score > 0.8](https://image.slidesharecdn.com/20230207datascienemilannexi-230216113550-2bd6a8b0/75/ML-Graph-algorithms-to-prevent-financial-crime-in-digital-payments-19-2048.jpg)
![This content is classified as Internal
18
send
money
Level 02 – Graph rule based approach
0.8 0.3
Org.
0.4
user 01 user 02
user 05
user 03
send
money
. . .
user N
send
money
0.93
0.2
user 06
0.3
• Network anomalies are hard to model and to check
with relational database
• Graph databases, like Neo4j, are the best choice when
your analytical focus is link between peers, that is a
network.
• Thanks to graph data structure several anomaly
detection rules can be built to fight groups of
suspicious transactions or users.
MATCH
(u:User)-[:send_money*]->(v:User)
WHERE
u.score > 0.8 and v.score > 0.8
Level 02 Description](https://crownmelresort.com/image.slidesharecdn.com/20230207datascienemilannexi-230216113550-2bd6a8b0/75/ML-Graph-algorithms-to-prevent-financial-crime-in-digital-payments-18-2048.jpg)
![This content is classified as Internal
19
Level 02 – Graph rule based approach
Data
Lake
ingest
AML
Feature
Store
Amazon
S3
Amazon
Glue
EC2
load
Graph
anomalies
Back-end
has
account
Front-end
query
process
DATA MODEL
1
Neo4j CE on EC2 (not managed)
2
Neo4J Spark connector
3
MATCH
(u:User)-[:send_money*]->(v:User)
WHERE
u.score > 0.8 and v.score > 0.8](https://crownmelresort.com/image.slidesharecdn.com/20230207datascienemilannexi-230216113550-2bd6a8b0/75/ML-Graph-algorithms-to-prevent-financial-crime-in-digital-payments-19-2048.jpg)
Uploaded byData Science Milan
ML & Graph algorithms to prevent financial crime in digital payments
This document discusses using machine learning and graph algorithms to prevent financial crime in digital payments. It presents a three level approach: Level 0 uses rule-based SQL queries to detect anomalies, Level 1 applies supervised machine learning to classify transactions, and Level 2 uses a graph database and rules to model network anomalies. Level 3 combines machine learning, graph algorithms, and personalized page rank to spread anomaly scores throughout a transaction network to identify suspicious groups. The strategies are being piloted through the Infinitech Project to develop technologies for applications in financial crime prevention, cybersecurity, and personalized products using AI, big data, IoT, and blockchain.
In this document
Powered by AI
Introduction to ML & Graph algorithms used by Nexi to combat financial crime in digital payments.
Nexi's end-to-end payment solutions for financial institutions, merchants, and consumers highlight customer-centric innovations.
Overview of Nexi's capabilities, workforce, IT spending, and market leadership in payment processing and merchant acquiring.
Current job openings at Nexi, including Data Scientist and Data Engineer positions.
Introduction to ML & Graph algorithms used by Nexi to combat financial crime in digital payments.
Agenda and objectives of the Infinitech Project, focusing on AI/IoT/Blockchain technologies in the fintech arena.
Identifies financial crime's impact on institutions, emphasizing money laundering, fraud, and the need for enhanced monitoring approaches.
Agenda and objectives of the Infinitech Project, focusing on AI/IoT/Blockchain technologies in the fintech arena.
Explores ML and Graph methodologies to improve anti-money laundering efficacy, detailing architecture and output rankings.
Agenda and objectives of the Infinitech Project, focusing on AI/IoT/Blockchain technologies in the fintech arena.
Links to Infinitech project resources, publications, and an invitation for questions.
Key takeaways from the presentation emphasize the importance of a hybrid rule-based and ML approach to AML.
Links to Infinitech project resources, publications, and an invitation for questions.
More Related Content
Similar to ML & Graph algorithms to prevent financial crime in digital payments
![[DSC Europe 22] Anti-Money Laundering ML Modeling approach - Gizem Akar](https://cdn.slidesharecdn.com/ss_thumbnails/gizemakarantimoneylaunderingmlmodel-221130080718-afd93383-thumbnail.jpg?width=640&height=640&fit=bounds)
![[Redis Released]- FalkorDB - Redis + Graph Agentic Memory’s Secret Sauce](https://cdn.slidesharecdn.com/ss_thumbnails/redisreleased-falkordbslidedeck-1125-251115194922-e1c0046b-thumbnail.jpg?width=640&height=640&fit=bounds)
ML & Graph algorithms to prevent financial crime in digital payments
- 1. This content isclassified as Internal ML & Graph algorithms to prevent financial crime in digital payments Alberto Danese, Head of Data Science Paolo Testa, Lead Data Scientist
- 2. This content isclassified as Internal 2 Next generation payments Next to Customers
- 3. This content isclassified as Internal 3 End-to-end payment solutions for Financial Institutions, Merchants and Consumers
- 4. This content isclassified as Internal 4 CAPABILITIES MANAGING SERVING SCALE >10.500 people >3.000 Product & Tech Development Specialists ~€300mln Annual Total IT & Innovation Spending ~170mln Cards ~2.2mln Merchants #1 Merchant acquirer by number of merchants and transaction value #1 Card processor by number of cards and transaction volume >1k Top financial Institutions ~15 bn ACH trx >25 Countries HQ The Leading PayTech, European by scale, Local by nature
- 5. This content isclassified as Internal 5 Select * from Nexi where dept like ‘DATA%’ ~40 internal people (in the hub - Data Area) including 10 in Data Science Team 5+ spokes within the business lines Support Business Development Fight Financial Crime Evolve Tech and Develop Special Projects 90% projects on public cloud Product-oriented mindset with end-to-end ownership
- 6. This content isclassified as Internal 6 Work with us! Check on Linkedin jobs for Nexi Italy, Nexi Group and Nexi Digital Currently, you may be interested in: • Junior Data Scientist – Fraud Intelligence Specialist (Nexi Italy) • Data Engineer (Nexi Digital)
- 7. This content isclassified as Internal 7 ML & Graph algorithms to prevent financial crime in digital payments
- 8. This content isclassified as Internal 8 INFINITECH PROJECT ML & GRAPH APPLICATION FINANCIAL CRIME IN DIGITAL PAYMENTS Agenda - ML & Graph algorithms to prevent financial crime in digital payments 01 What are the main challenges related fo financial crime? What is Anti- money laundering (AML)? 02 How can ML and Graph help AML ot optimize and improove their daily activity? 03 What is the Infinitech Project and you can access to it?
- 9. This content isclassified as Internal 9 Financial crime challenges • Financial crime is one of the main challenges that banks, and financial institutions must face, leading to : • Financial penalties • Reputational Damage • Lose customer trust • While security measure are more and more advanced, fraudsters and organized crime keep on evolving as well • Two types of criminal activity: • Payments Fraud → the cardholder is victim of fraudulent organization. • Money Laundering → the cardholder is involved in criminal activity, using payments tools. In 2020 global banks were hit with $10.4bn in fines for money-laundering violations, an increase of more than 80% on 2019
- 10. This content isclassified as Internal 10 Anti-Money Laundering (AML) • Money Laundering happens when individuals attempt to hide profit gained from criminal means • Money can be originated by gambling, drug trafficking, and other illegal activities • Nexi AML aim is to put risk procedures to recognize and stop criminal activities using Nexi products (cards or POS). Money laundering in digital payments AML Transaction Monitoring workflow 1. Suspicious payments transaction Alert is generated 2. The AML investigators analyze the data related to the alert 3. The investigator evaluates whether to notify to central authority (Bank of Italy informative financial unit) the suspicious case
- 11. This content isclassified as Internal 11 Anti-Money Laundering Challenges “Human in the Loop” is a must We can’t fully automize and optimize the transaction monitoring, since evaluation of AML investigator is a requirement. But we can optimize the volume of alerts, that is, the suspicious transactions to analyze # of analysis of payments transactions time & cost effectiveness FP – FN trade off FP Risk of sanctions due to missed notification You can’t analyze each single transactions FN FP FN
- 12. This content isclassified as Internal 12 INFINITECH PROJECT ML & GRAPH APPLICATION FINANCIAL CRIME IN DIGITAL PAYMENTS Agenda - ML & Graph algorithms to prevent financial crime in digital payments 01 What are the main challenges related fo financial crime? What is Anti- money laundering (AML)? 02 How can ML and Graph help AML ot optimize and improove their daily activity? 03 What is the Infinitech Project and you can access to it?
- 13. This content isclassified as Internal 13 Level 0 – rule base approach Data Lake Amazon S3 SQL-based Anomaly Engine Amazon Athena Reporting Layer Microsoft PowerBI SELECT user_id FROM transactions WHERE metric_1 > THR1 and metric_2 > THR2 PROS Simplicity: each alert is triggered by an SQL Explainability : a SQL is rule shared with AML CONS Scale: high transact. volume, +300 rules, generating large amount of alerts, leading to high False Positive rate Group: relational dbs are not the best solutions for network modelling Highlights Architecture
- 14. This content isclassified as Internal 14 Level 01 - Supervised ML The first approach is to train a ML model to classify anomalous cases based on notification pattern (target) Thanks to this, we avoid to set manually +500 rules; we just metrics of those rule as features in the ML model. The output of the model will be probability score of a customer to be notified due to suspicious transactions AML investigators can focus only on tuning only 1 threshold (the prob. Score) and can be used as rank metric User id Feature 01 Feature 02 … Feature P Target 373429 12.000 € 23% 1 1 598492 1.000 € 10% 0 0 … … … .. .. .. 598492 160.000 € 37% 1 0 Level 01 Description
- 15. This content isclassified as Internal 15 Level 01 – Deep dive Data Lake Amazon S3 process AML Feature Store prepare train test score datasets Amazon SageMaker train Score & SHAP ML score Back-end Front-end report Pipeline Orchestration AWS Step Functions SERVERLESS & FULLY MANAGED STACK Amazon Glue Model Registry Experiment Registry 1 2 3 Serverless & Fully Managed ML architecture AML FS is part of Nexi FS. Great_expectations and TDD guarantees us data quality 4 Data Centric AI SageMaker transaparently store train jobs artifacts and experiemnts in dedicated registry
- 16. This content isclassified as Internal 16 Level 01 – Highlights ML score Shap values ordered features Top 10000 risk Notified
- 17. This content isclassified as Internal 17 Level 01 – Highlights • The output is a probability score; 1 threshold to fine tune the results • SHAP values guarantees explainability PROS CONS • The score is associated to a single independent user, NOT in a Network • The ML model learn only from notification of AML investigators.
- 18. This content isclassified as Internal 18 send money Level 02 – Graph rule based approach 0.8 0.3 Org. 0.4 user 01 user 02 user 05 user 03 send money . . . user N send money 0.93 0.2 user 06 0.3 • Network anomalies are hard to model and to check with relational database • Graph databases, like Neo4j, are the best choice when your analytical focus is link between peers, that is a network. • Thanks to graph data structure several anomaly detection rules can be built to fight groups of suspicious transactions or users. MATCH (u:User)-[:send_money*]->(v:User) WHERE u.score > 0.8 and v.score > 0.8 Level 02 Description
- 19. This content isclassified as Internal 19 Level 02 – Graph rule based approach Data Lake ingest AML Feature Store Amazon S3 Amazon Glue EC2 load Graph anomalies Back-end has account Front-end query process DATA MODEL 1 Neo4j CE on EC2 (not managed) 2 Neo4J Spark connector 3 MATCH (u:User)-[:send_money*]->(v:User) WHERE u.score > 0.8 and v.score > 0.8
- 20. This content isclassified as Internal 20 Level 02 – Graph rule based approach • Now network anomaly scenarios are now addressed • Graph model are easy to explain and straightforward in rule settings. PROS CONS • We lose the scale factor given by automation of ML model. It’s a manual rule based apporach made with Cypher queries, instead of SQL
- 21. This content isclassified as Internal 21 Level 03 – ML and Graph USER ML SCORE RANK user 03 0.93 1 user 01 0.8 2 user 05 0.4 3 user 02 0.3 4 user 04 0.2 5 ML anomaly only consider single users. If a user within a network with high risk, then we want to spread this information to this user. USER ML SCORE NEW RANK ML RANK user 03 0.93 1 1 user 01 0.8 2 2 user 02 0.3 3 (+1) 4 user 04 0.2 4 (+1) 5 user 05 0.4 5 (-2) 3 What we need is an algorithm that spread ML score information to within a given network topology to adjust score and find group of anomalous case. Level 03 Description
- 22. This content isclassified as Internal 22 Page Rank (PR) is an iterative algorithm originally used by Google to rank web pages (nodes) according to the quality and quantity of links (edges) pointing to the pages. Personalized Page Rank 𝑥′ = (1 − α) 𝑃′𝑥 + α ν 𝑥 = importance rank ν = intrinsic centrality P = adjacency matrix We use ML score as intrinsic centrality (v) to spread anomaly information within the network, with the aim to find groups of suspicious individuals. personalized PR page-rank
- 23. This content isclassified as Internal 23 Level 03 – Architecture & Contribution Data Lake AML Feature Store ML anoamly Score ingest PPR Back-end load Front-end Apply PPR 1 Level01 ML anomaly score is loaded into Graph db as node attribute 2 Customized PPR is applied as a Cypher function over the graph database 3 PowerBI report show PPR score
- 24. This content isclassified as Internal 24 Level 03 - Highlights • Scale factor is addressed, since PPR run is operationalized. • The PPR rank combines the best of Level 01 (ML) and Level 02 (graph algorithms) approaches • As for Level 02, graph and PPR is highly interpretable and explainable PPR RANK ML RANK ML SCORE 3 4859 0.93 3 neighbors
- 25. This content isclassified as Internal 25 Summary Level 0 – SQL rule based approach Level 1 – Supervised ML Level 2 – Graph rule based approach Level 3 – ML and Graph based approach INFINITECH INFINITECH
- 26. This content isclassified as Internal 26 INFINITECH PROJECT Flagship initiative in Digital Finance within ICT Horizon 2020 Programme Project deliverables: • Marketplace of AI/IoT/Blockchain/Cybersecurity technologies (algorithms, real-time analytics, blockchain based data sharing) • Testbeds and sandbox for fintech players to test and deploy their own solutions • Regulatory tools to helps company to facilitate compliance to regulations
- 27. This content isclassified as Internal 27 INFINITECH PILOTS Nexi Payments is one among 15 organizations, commercial banks, central banks, startup, insurance firms, who joined Infinitech project as pilot. Fintech/Insuretech application fields: • Financial crime • Cybersecurity • Personalized products Technologies: • #AI • #BigData • #IoT • #Blockchain
- 28. This content isclassified as Internal 28 INFINITECH OUTPUT https://pilot16.infinitech-h2020.eu/ https://gitlab.infinitech-h2020.eu/pilot16/ aml-graph-payments-anomaly-detection Two are outputs in the Infinitech Testbed/Sandbox: 1. PILOT DEMO WEB-APP Web app to show algorithms ( both anomaly detection and PPR) with third part or synthetic generated data 2. ANOMALY ENGINE w/ JUPYTER Git repository with docker containers to use pilot graph algorithms on third-part data in a Jupyter notebook environment
- 29. This content isclassified as Internal 29 1. PILOT DEMO WEB-APP https://pilot16.infinitech-h2020.eu/ Web app to show algorithms ( both anomaly detection and PPR) with third part or synthetic generated data
- 30. This content isclassified as Internal 30 2. ANOMALY ENGINE w/ JUPYYER Git repository with docker containers to use pilot graph algorithms on third-part data in a Jupyter notebook environment https://gitlab.infinitech-h2020.eu/pilot16/aml-graph-payments-anomaly-detection
- 31. This content isclassified as Internal 31 Pubblications and Contributions
- 32. This content isclassified as Internal 32 Wrap up and special thanks What did we learn along the way? • Rule based approaches are still present and widely adopted in 2023: that’s fine! But ML is a sound approach when dealing with hundreds of rules • Human in the loop is key: optimizing the work of an analyst may be even better than a fully automated approach (in some scenarios) • Network is a natural representation for some phenomena, like digital payments: a good data scientist should be able to find the best combination of ML & graph data science • EU projects are a great opportunity!
- 33. This content isclassified as Internal 33 Wrap up and special thanks It’s been a team-work with multiple people giving support in a way of another! Thank you to: • Fabio Dezi, Luca Latella and Luca Rinaldi – and the always supporting DST • Alfredo Fomitchenko and Vittorio Giatti • Laura Arditti, Alberto De Lazzari and the LARUS team • Infinitech organization and GFT SPECIAL THANK YOU TO… What did we learn along the way? • Rule based approaches are still present and widely adopted in 2023: that’s fine! But ML is a sound approach when dealing with hundreds of rules • Human in the loop is key: optimizing the work of an analyst may be even better than a fully automated approach (in some scenarios) • Network is a natural representation for some phenomena, like digital payments: a good data scientist should be able to find the best combination of ML & graph data science • EU projects are a great opportunity!
- 34. This content isclassified as Internal 34 ANY QUESTION?
- 35. This content isclassified as Internal