AML Transaction Detection

Exploratory data analysis and machine-learning pipeline for the IBM Transactions for Anti-Money Laundering (AML) synthetic dataset.

Dataset: https://www.kaggle.com/datasets/ealtman2019/ibm-transactions-for-anti-money-laundering-aml

Quick start

# 1. Install uv (if not present)
curl -LsSf https://astral.sh/uv/install.sh | sh

# 2. Create virtual environment and install dependencies
uv sync

# 3. Download dataset (requires Kaggle API token at ~/.kaggle/kaggle.json)
uv run python -c "
import kaggle
kaggle.api.authenticate()
kaggle.api.dataset_download_files(
    'ealtman2019/ibm-transactions-for-anti-money-laundering-aml',
    path='data/raw', unzip=True
)
"

# 4. Launch Jupyter
uv run jupyter lab

Dataset

The benchmark family includes multiple HI / LI and Small / Medium / Large splits, but this repo currently analyses only the HI-Large split.

• Notebook transaction input: data/raw/HI-Large_Trans.csv
• Notebook patterns input: data/raw/HI-Large_Patterns.txt
• Other splits and companion files are not used by notebooks/01_eda.ipynb

In the broader benchmark:

• HI = high-illicit-ratio split, with roughly 5% laundering-labelled transactions
• LI = low-illicit-ratio split, with roughly 0.1% laundering-labelled transactions
• Small / Medium / Large refer to dataset size, not a different schema

Group	Size	Approx illicit ratio
HI	Small	~5 %
HI	Medium	~5 %
HI	Large	~5 %
LI	Small	~0.1 %
LI	Medium	~0.1 %
LI	Large	~0.1 %

Each split ships with three companion files:

• *_Trans.csv = transaction-level records
• *_accounts.csv = account metadata
• *_Patterns.txt = ground-truth laundering pattern blocks

Columns

Column	Description
Timestamp	Date-time of the transaction
From Bank	Originating bank ID
Account (from)	Originating account ID
To Bank	Receiving bank ID
Account (to)	Receiving account ID
Amount Received	Amount received (in Receiving Currency)
Receiving Currency	ISO currency code at destination
Amount Paid	Amount paid (in Payment Currency)
Payment Currency	ISO currency code at source
Payment Format	Wire, Cheque, Credit Card, ACH, etc.
Is Laundering	Binary label – 1 = laundering, 0 = legitimate

Laundering patterns

• Fan-out – One account rapidly sends to many recipients
• Fan-in – Many accounts consolidate to one
• Bipartite – Many-to-many transfer block between sender and receiver sets
• Cycle – Money circulates through a closed loop of accounts
• Gather-Scatter – Aggregate then disperse through layering
• Scatter-Gather – Disperse then re-aggregate
• Stack – Layered pass-through chains
• Random – Irregular mixing pattern

Example

The structuring / smurfing view from the EDA is one of the clearest visuals in the notebook:

Project layout

aml-transaction-detection/
├── data/
│   ├── raw/          <- original Kaggle CSVs
│   └── processed/    <- parquet caches
├── notebooks/
│   └── 01_eda.ipynb  <- main EDA notebook
├── src/              <- helper modules (future)
├── pyproject.toml
└── README.md