What a Bank's Fraud Graph Teaches You About Chargebacks

In February, researchers from LaBRI (University of Bordeaux) published a study with BNP Paribas Personal Finance on graph-based fraud detection. The bank reports a 20% fraud reduction on its installment payment product. The technique runs on order history, and you already own that data. That is why the study deserves a merchant's attention.

The principle fits in one sentence. Every order becomes a node on a graph. Two orders get linked when they share customer information: same email, same address, same phone. An honest customer shows up alone or in a small cluster. A fraudster who juggles identities leaves a pattern behind, separate orders pointing to the same set of details.

What the study measured

The team tested the idea on 500,000 real orders with a 3.49% fraud rate. Three results:

1. Removing topology indicators drops detection precision by 27 to 31%. The graph is a signal in its own right.
2. Combining graph embeddings (GraphSAGE) with those indicators improves precision by 10% over the baseline.
3. The most precise model is the hardest to explain. In France, a refused customer can legally demand the reasons for a refusal, so explainability is part of the job.

Fraud rings hit merchants too

Banks are one target. Merchants are the other. A ring places orders with slightly different identities, then the chargebacks arrive in series. Most get filed under fraud codes; for Visa that is usually 10.4, Other Fraud: Card-Absent Environment.

The same graph logic works on your chargeback data. If five chargebacks over three months share the same normalized email (dots and Gmail aliases included), the same IP address, or the same device, that is not five angry customers. It is one pattern.

Three signals worth checking in your data

• The same email across several chargebacks, normalized: dots and Gmail aliases included
• The same shipping address or the same device fingerprint
• Chargebacks arriving in waves, a few days apart

What the gap looks like in numbers

If your current stack reviews each chargeback in isolation, you see five separate problems where there is one fraudster. The study puts a number on the cost. Remove the graph signals and precision falls 27 to 31%. Add graph embeddings on top of classic indicators and it climbs 10%.

Read the details yourself

The dataset behind the study is public on Zenodo (doi 10.5281/zenodo.14216924). The methodology is in the paper on HAL (hal-04939824), a collaboration between LaBRI and BNP Paribas Personal Finance.

Linking chargebacks by hand across a spreadsheet stops scaling fast. ChargeGuard links them automatically, drafts the response for each reason code, and its free audit shows what your data looks like as a graph. We also keep plain-English guides for the nine reason codes merchants hit most.