Fraud and financial crime

Graph analytics for fraud rings and AML networks: finding what row-by-row models miss

Fraud rings and money-laundering networks hide in relationships: shared devices, beneficiaries, addresses and fast chains of transfers. DataNature builds an entity graph from resolved customers, accounts, devices and transactions in the lakehouse, adds edges in real time from StreamHouse, and uses graph features and traversals to flag rings, mule networks and layering, while Voir agents help investigators follow and document each path.

Published . Last updated

On this page

Most fraud and anti-money-laundering (AML) models score one transaction or one customer at a time. Organised fraud is designed to look normal at that level: each account is new but plausible, each transfer is under a threshold. The signal is in the network. This guide explains how to build graph analytics for fraud and AML on DataNature, from entity resolution to investigator workflows.

Why do fraud rings and laundering networks need graph analytics?

Because their defining features are relationships. A mule network is many accounts receiving funds from victims and forwarding them quickly to a few collectors. A synthetic-identity ring is many applicants sharing phone numbers, addresses or devices. Layering is a chain of fast transfers across accounts and institutions. Row-level models see each piece as low risk; a graph shows the shape.

What does the graph contain?

Node typeExamples of edges
Customer (resolved identity)owns account, shares address, shares phone, same employer
Accountsends to, receives from, linked card, opened via device
Device and IPused by account, used by customer, seen at login
Beneficiary and merchantpaid by account, refunds to account
Address, phone, emailregistered by customer, used in application
Case and alertinvolves customer, involves account, closed as fraud

Entity resolution in Silver comes first. If the same person appears under three customer records, the graph would show three weak nodes instead of one strong one. dbt models match records on identifiers and normalised attributes, and keep the match rules versioned.

Which graph signals matter?

  • Shared attributes: many customers on one device, phone or address, especially new ones.
  • Proximity to known bad: accounts within one or two hops of confirmed fraud or a closed mule account.
  • Fan-in, fan-out: many senders into one account followed by fast transfers out.
  • Velocity along paths: funds moving through three or more accounts within hours.
  • Communities: tightly connected groups of accounts that transact mostly with each other.
  • Circular flows: money returning to its origin through intermediaries.

These become graph features (degree, hop distance to a confirmed case, community size, share of new accounts in the community) that feed the same fraud and AML models as transactional features.

How do real-time and batch work together?

  • StreamHouse adds edges as events happen: a new device registration, a new beneficiary, a transfer. Real-time scoring can then ask, while the transfer is in flight, "is the destination within two hops of a known mule?"
  • Batch graph jobs on Spark recompute heavier features nightly over the full history: communities, path velocities, risk propagation.
  • Iceberg holds both, so models train on the exact graph features that were available at decision time.

How do investigators use it?

Alerts are only useful if investigators can understand them quickly. Voir twins for the financial-crime team:

  1. Expand the network around an alerted account, within the investigator's scope.
  2. Summarise flows, shared attributes and related cases, pulling figures from governed cubes.
  3. Pull supporting documents such as KYC files and prior case notes from DocuContext.
  4. Draft the case narrative or suspicious activity report with citations to every fact.
  5. Wait for approval before filing or escalating; every step is logged.

What controls does this need?

  • Customer identifiers hashed for analysts, with investigators exempt.
  • Row filters by legal entity or country where data-sharing rules apply.
  • Lineage from each graph feature to its source tables, for model risk review.
  • Decisions, features and model versions written back to Iceberg for audit.
  • Everything inside the institution's jurisdiction, including the graph and agent logs.

Frequently asked questions

Do we need a separate graph database?

DataNature builds and governs the entity graph on the platform, from lakehouse tables and StreamHouse events, so investigators and models use the same governed data without exporting it to another system.

How do graph features improve models?

They add network context that row-level features cannot capture, such as distance to confirmed fraud or membership in a fast-growing community of new accounts. They are used alongside, not instead of, transactional features.

Can this support suspicious activity reporting?

Yes. Voir can draft narratives with citations for an investigator to review and approve, and every fact links back to its source.

Sources

How does DataNature Atlas keep this secure and compliant?

Investigators see what they need; analysts see masked identifiers; every score and decision is auditable. DataNature Atlas is the platform's governance module:

  • Knows your data: a catalog with automated PII detection and classification across Bronze, Silver and Gold.
  • Controls access: role-based grants plus attribute-based policies, row filters and column masks (null, redact, partial, hash, year only).
  • Governs AI: Voir agents are scoped principals; masks, row filters and PII guardrails apply to every tool call and answer.
  • Proves it: lineage from source to metric to decision, and an audit log of who or what accessed which data.
  • Stays local: Atlas runs inside your deployment, so policies, lineage and logs never leave your jurisdiction.

These controls support GDPR, PIPEDA, Quebec Law 25, CCPA, HIPAA and sector rules such as OSFI E-23. Read DataNature Atlas in depth or see security and compliance.

See it on your own data.

Book a demo

Demos and quotes are handled by Dlytica, the company behind DataNature.