gen_fraud_graph

Synthetic fraud graph generator for training and benchmarking graph-based fraud detection models in financial services.

---

Overview

gen_fraud_graph is an open-source Python tool that generates massive synthetic financial transaction graphs with injected fraud patterns and optional vector embeddings. It produces CSV datasets ready for ingestion into graph databases (TigerGraph, Neptune, Neo4j, JanusGraph) or for training graph neural networks (GNN).

The generator creates three types of data: - Account nodes — synthetic customer accounts with balance, risk score, and optional embedding vectors - Transaction edges — normal financial transactions between accounts - Fraud rings — cyclic money-laundering patterns with suspicious transaction descriptions

Key Features

• Massive scale — Generate from 1K to 100M+ accounts with configurable scale factor
• Fraud pattern injection — Cyclic money-laundering rings with configurable depth (4–7 hops)
• Parallel generation — Multi-process workers for fast generation on high-core machines
• Vector embeddings — Three providers: fake (random, fast), local (SentenceTransformers), openai (API)
• Multiple formats — Generic CSV or AWS Neptune bulk-load format
• Resume support — Interrupted generation can resume from where it left off
• Privacy by design — All data is 100% synthetic; no real financial data is used

Use Cases

• Training and evaluating graph neural networks (GNN) for fraud detection
• Benchmarking anti-money laundering (AML) detection algorithms
• Load-testing graph databases (TigerGraph, Neptune, JanusGraph, NebulaGraph, FalkorDB)
• Research in financial crime detection and anomaly detection on graphs
• Generating labeled datasets for deep learning on graph-structured data

---

Quick Start

Installation

Note: gen-fraud-graph is not yet published to PyPI, so pip install gen-fraud-graph will fail with No matching distribution found. Until the first PyPI release, install from source as shown below. The PyPI badge above is pre-provisioned for the planned release.

Install from source using uv:

git clone https://github.com/SantanderAI/gen-fraud-graph.git
cd gen-fraud-graph
uv venv && source .venv/bin/activate
uv pip install -e '.[dev]'

With optional embedding providers (from the cloned source directory):

uv pip install -e '.[local]'    # SentenceTransformers (local model)
uv pip install -e '.[openai]'   # OpenAI API embeddings
uv pip install -e '.[all]'      # Everything including dev tools

If you prefer plain pip over uv, the source install works the same way:

python -m venv .venv && source .venv/bin/activate
pip install -e '.[dev]'

Once the package is published, pip install gen-fraud-graph will be the recommended path.

CLI Usage

# Quick test (~1K accounts, ~9K transactions, fake embeddings)
gen-fraud-graph --scale 0.0001 --provider fake --output ./data

# Medium scale (~100K accounts, parallelized)
gen-fraud-graph --scale 0.01 --workers 4 --output ./data

# Full benchmark (~10M accounts, ~90M transactions)
gen-fraud-graph --scale 1.0 --workers 24 --output ./data

# Neptune bulk-load format
gen-fraud-graph --scale 0.01 --format neptune --output ./neptune_data

# Resume interrupted generation (skips completed files)
gen-fraud-graph --scale 1.0 --workers 24 --skip-accounts --output ./data

CLI Arguments

Flag	Default	Description
--scale	1.0	Scale factor. 1.0 = ~10M accounts / ~90M transactions. 0.01 = ~100K accounts.
--provider	fake	Embedding provider: fake (random vectors), local (SentenceTransformers), openai.
--output	data	Output directory for generated CSV files.
--workers	1	Number of parallel worker processes.
--batches	1	Number of file chunks per worker.
--format	csv	Output format: csv (generic) or neptune (AWS Neptune bulk-load).
--fraud-rings	auto	Number of fraud rings. Default: auto-scaled from --scale.
--compress	off	ZIP-compress output CSV files.
--skip-accounts	off	Skip account generation (useful when resuming).

Python API

# Copyright (c) 2026 Santander Group
# SPDX-License-Identifier: Apache-2.0

from gen_fraud_graph import Config, FraudGraphGenerator

config = Config(
    scale_factor=0.001,         # ~10K accounts, ~90K transactions
    num_fraud_rings=50,         # 50 cyclic fraud patterns
    embedding_provider="fake",  # random vectors (fast, no model needed)
    workers=2,                  # 2 parallel processes
    output_dir="./output",
)

generator = FraudGraphGenerator(config)
generator.run()

Verify Generated Patterns

python -m gen_fraud_graph.verify --data-dir ./data

---

Output Structure

data/
├── accounts/
│   ├── accounts_0_0.csv       # Account nodes (worker 0, batch 0)
│   └── accounts_1_0.csv       # Account nodes (worker 1, batch 0)
├── transactions/
│   ├── transactions_0_0.csv   # Transaction edges (worker 0, batch 0)
│   └── transactions_1_0.csv   # Transaction edges (worker 1, batch 0)
└── fraud/
    ├── transactions_fraud.csv  # Fraud ring transaction edges
    └── fraud_cases.csv         # Fraud ring metadata (pattern_id, accounts, depth)

CSV Schema

accounts (accounts_*.csv)

Column	Type	Description
account_id	string	Unique account identifier (acc_0, acc_1, ...)
customer_name	string	Synthetic customer name
balance	float	Account balance (100 – 100,000)
risk_score	float	Risk score (0.0 – 1.0)
creation_date	string	Account creation date

transactions (transactions_*.csv)

Column	Type	Description
tx_id	string	Unique transaction identifier
src_id	string	Source account
dst_id	string	Destination account
amount	float	Transaction amount (10 – 500 for normal, 9999 for fraud)
timestamp	string	Transaction timestamp
description	string	Transaction description
embedding	string	Pipe-separated embedding vector

fraud_cases (fraud/fraud_cases.csv)

Column	Type	Description
pattern_id	string	Pattern identifier (pat_0, pat_1, ...)
start_acc_id	string	First account in the ring
pattern_type	string	Always "cycle"
depth	int	Number of hops in the ring (4–7)
involved_accounts	string	Pipe-separated list of accounts

---

Scale Reference

Scale	Accounts	Transactions	Fraud Rings	Approx. Size
0.0001	1,000	9,000	10	~2 MB
0.001	10,000	90,000	10	~20 MB
0.01	100,000	900,000	10	~200 MB
0.1	1,000,000	9,000,000	100	~2 GB
1.0	10,000,000	90,000,000	1,000	~20 GB

---

Project Structure

gen_fraud_graph/
├── src/gen_fraud_graph/
│   ├── __init__.py       # Package entry point
│   ├── cli.py            # CLI (gen-fraud-graph command)
│   ├── config.py         # Configuration dataclass
│   ├── embeddings.py     # Embedding providers (fake/local/openai)
│   ├── exporters.py      # CSV/ZIP output writers
│   ├── generator.py      # Core 3-phase pipeline orchestrator
│   ├── typologies.py     # Fraud ring generator
│   └── verify.py         # Pattern verification utility
├── tests/
│   └── test_generator.py # Unit and integration tests
├── examples/
│   └── basic_usage.py    # Minimal Python API example
├── .github/
│   ├── workflows/        # CI (ci, codeql, dep-scan, license-check,
│   │                     #     pattern-check, cla, stale, release)
│   ├── ISSUE_TEMPLATE/   # Bug + feature templates
│   ├── PULL_REQUEST_TEMPLATE.md
│   ├── dependabot.yml    # Weekly Python + Actions updates
│   └── pattern-check-allowlist.txt
├── pyproject.toml        # Package metadata and tool config
├── LICENSE               # Apache 2.0
├── NOTICE                # Apache 2.0 attribution
├── CONTRIBUTING.md       # Contribution guidelines
├── CODE_OF_CONDUCT.md    # Contributor Covenant v2.1
├── SECURITY.md           # Vulnerability disclosure policy
├── CODEOWNERS            # Maintainer approvals
└── CHANGELOG.md          # Release history

---

Requirements

Core (always installed): - Python >= 3.10 - NumPy >= 1.24 - Pandas >= 2.0 - tqdm >= 4.65

Optional: - sentence-transformers >= 2.2 — for --provider local - openai >= 1.0 — for --provider openai

---

Contributing

We welcome contributions from the community. Please read our CONTRIBUTING.md before submitting a pull request.

By contributing, you agree to the terms of our Contributor License Agreement (CLA).

---

Security

To report a security vulnerability, please follow the process described in SECURITY.md. Do not open a public issue for security vulnerabilities.

---

License

This project is licensed under the Apache License 2.0 — see the LICENSE file for details.

Copyright (c) 2026 Santander Group
SPDX-License-Identifier: Apache-2.0

---

Citation

If you use this tool in your research, please cite:

@software{gen_fraud_graph,
  title     = {gen\_fraud\_graph: Synthetic Fraud Graph Generator},
  author    = {Santander AI Lab},
  year      = {2026},
  url       = {https://github.com/SantanderAI/gen-fraud-graph},
  license   = {Apache-2.0}
}

---