9bNine Bits

Understanding FDS with Graph Modeling

In a fraud detection system (FDS), graph modeling represents accounts, devices, cards, and transactions as connected nodes instead of normalized tables. This turns multi-hop relationships, such as several accounts funneling money through one shared device, into a direct traversal instead of a chain of joins. The same approach extends past finance to cyber threat intelligence, where indicators of compromise and infrastructure are linked the same way.
Understanding FDS with Graph Modeling

Understanding Graph Modeling

Before understanding FDS with graph modeling, there are three primary things to go over when considering graph modeling for analytical purposes.

 

First, an analysis of data and domain must be the basis of understanding. Based on this, it is necessary to investigate which analysis services and systems have been studied and researched at the project’s site. When the data and domain research is complete, review if there are any issues or limitations within the area of the current domain. This is a necessary step to see if designing graph modeling can provide value. Modeling should only be considered if a difference between existing services and graph modeling is worth exploring.

 

After data and domain analysis, the next step of the modeling research and verification process has to be carried out. The purpose of this step is to examine the research and hypotheses of the graph model while looking at experimental designs from various angles.

 

The final step is deriving value and verifying the result via the graph model and internalizing it. Verifying the selected graph model and internalizing the value is necessary when applying it to the actual service and performing analysis on similar domains.

 

Next, let’s look at how FDS (Fraud Detection System) using modeling applies to the graph modeling steps above.

 

FDS Modeling from an Analytical Perspective

The existing FDS detects fraud from accumulated data in real-time based on predefined rules. If a graph is to be implemented here, the most basic approach would be to connect all transactions and actions using the ‘connection’ characteristic of the graph.

 

There are two main modeling methods that are often used as a combination in FDS. The first is to model the transaction behavior and flow as a single graph. Abnormal transactions and fraudulent activities leave some type of traces. Modeling information of time and transaction from these traces creates a new model that will be able to detect transaction amount, the flow of the transaction, and the relationship between abnormal accounts.

 

Transactions between accounts via single graph modeling

 

The second method is using heterogeneous network modeling, which is practically adding and connecting data such as the personal information of the analysis target. The heterogeneous graph modeling is able to derive complex relationships that proved to be difficult to find in a single graph. The relationship acts as a bridge between the nodes and the data that plays the role of that bridge becomes common attribute information, as shown in the image below.

 

An example of heterogeneous graph modeling

 

 

Summary: 4 Values of FDS with Graph Modeling

The following four core values summarize the benefits of implementing graph modeling in FDS.

 

  1. Deriving transaction patterns: Graph modeling is able to define and derive a pattern from transaction flow and the transaction relationship with an abnormal account.
  2. Searching data for fraudulent behavior: Modeling both the transactional and fraudulent behaviors into nodes and edges of the graph can help users to understand data structure more intuitively and efficiently.
  3. Circular loop pattern: The circular loop pattern is generated while modeling a heterogeneous graph. When a relationship is established across the ‘bridge’, the data structure is stored in the form of a circular loop consisting of nodes and edges. Such a pattern is also known as the ‘Fraud Ring’ in the FDS domain. The fraud ring’s strength lies in deriving patterns from graph models.
  4. Graph projection: A relationship established in heterogeneous graph modeling is a relationship that does not exist in actual raw data. Graph projection allows converting of the relationship from a heterogeneous graph model into a single graph model. The advantage of graph projection is that graph analysis can be executed on new relationships that are formed during the conversion to a single graph.

 

 

 

 

 

Related

Leave a Reply Cancel reply

Your email address will not be published.

Save my name, email, and website in this browser for the next time I comment.

Notify me of follow-up comments by email.

Notify me of new posts by email.

Common questions

What does FDS mean in the context of graph modeling?

FDS stands for fraud detection system: the scoring logic and checks that flag transactions, accounts, or sessions as high risk for fraud. Modeling the underlying entities as a graph makes connected or circular activity easier to query than it would be in flat tables.

Can graph modeling be done directly in PostgreSQL?

Yes. PostgreSQL extensions such as Apache AGE add a property-graph type and a pattern-matching query language on top of standard PostgreSQL storage, so a graph model can live in the same database as relational tables.

What is a graph projection used for?

A graph projection is a narrower copy of a graph, limited to specific node and edge types, built so an algorithm like community detection or shortest-path scoring can run against only the data it needs instead of the entire operational graph.

How is a circular loop pattern detected in a fraud graph?

It is detected with a path query that matches a chain of directed edges returning to its starting node, usually bounded by a maximum path length and filtered by time window and transaction amount so an incidental cycle is not mistaken for a real layering pattern.

Should fraud detection graphs use row storage or columnar storage?

It depends on the workload. Row storage suits the frequent writes of new transactions and sessions, while columnar or in-memory storage suits large analytical scans and scoring jobs, which is why many teams keep the live graph in row storage and periodically materialize a separate analytical projection.

Sources

  1. PostgreSQL Documentation
  2. Apache AGE Project
  3. openCypher Project

About this page

Last reviewed October 30, 2026 by the Nine Bits editors. Read our editorial policy or report an error.