Question about fraud signal design and GNN applicability in generated dataset

[mauuuuu5]

Hi team,
First of all, thank you for open-sourcing this project. I've been working with gen-fraud-graph to understand how synthetic fraud data can be used for GNN-based detection, and I have two questions about the design choices.

1. Fraud signal is trivially detectable by non-structural attributes
   After analyzing the generated data, I found that fraudulent transactions are perfectly separable using only three attributes:

Attribute	Fraud	Normal
Amount	$9,999 fixed	$10–$500
Description	"rapid movement", "structuring", "shell company"	"utility bill", "grocery", "salary"
Timestamp	12:00:00 fixed	Distributed across hours

Question: Was this level of separability intentional for pedagogical purposes, or should the generator create more realistic fraud patterns where structural graph features (e.g., cycles, centrality) are necessary for detection?

2. Cycles exist for both fraud and normal transactions
   I also analyzed whether the cyclic structure distinguishes fraud. Using networkx.simple_cycles(length_bound=7):

Metric	Fraud	Normal
Transactions in cycles	100%	100%
Average cycles per edge	~1,053	~1,054

Every edge participates in ~1,000 cycles, meaning the graph is extremely dense. The README states the fraud is "cyclic money-laundering rings with configurable depth (4–7 hops)," but this structure appears to be a property of the entire graph, not just fraud.

Question: If the cyclic structure is not discriminative, what is the intended graph-based signal that a GNN should learn? Is the generator designed primarily for database load testing (as mentioned in the README) rather than ML benchmarking?

The dataset generated was

!gen-fraud-graph \ --scale 0.0001 \ --provider fake \ --fraud-rings 180 \ --output data