README.md

---
license: cc-by-4.0
task_categories:
- graph-ml
- tabular-classification
language:
- en
tags:
- anti-money-laundering
- AML
- financial-crime
- fraud-detection
- graph-neural-networks
- GNN
- synthetic
- banking
- transactions
- PyTorch-Geometric
- Neo4j
pretty_name: AusAML — Synthetic Australian Banking Dataset for AML Detection
size_categories:
- 10M<n<100M
---

# AusAML: Synthetic Australian Banking Dataset for AML Detection

AusAML is a large-scale synthetic banking dataset designed for training and benchmarking graph neural network (GNN) models on anti-money laundering (AML) detection tasks. It simulates a realistic four-bank Australian financial ecosystem with fully labelled AML typology scenarios embedded within a background of legitimate transaction activity.

## Dataset Summary

| Property | Value |
|---|---|
| Jurisdiction | Australia (AU) |
| Banks | 4 (Yellow, Red, White, Orange) |
| Customers | 50,000 (primary) |
| Accounts | 112,620 |
| Transactions | 35,554,888 |
| Date range | 2025-11-01 → 2026-05-31 (7 months) |
| AML customers | 2,502 (5.0% of total) |
| AML scenario instances | 740 across 4 banks (297 unique) |
| AML typologies | 29 |
| Random seed | 42 |

All personally identifiable information is entirely synthetic. No real customer, account, or transaction data was used.

---

## AML Typologies

The dataset covers 29 distinct money-laundering typologies spanning cash structuring, network-based layering, digital-asset obfuscation, and professional abuse patterns:

| Category | Typologies |
|---|---|
| Cash structuring | `smurfing`, `micro_structuring`, `exchange_micro_structuring`, `geo_smurfing` |
| Network layering | `layering`, `chain_branch`, `circular_ring`, `split_merge`, `fanin_fanout`, `hub_network`, `mule_network`, `pass_through`, `funnel_account`, `intrabank_hub` |
| Burst & timing | `burst_transfers`, `dormant_activation` |
| Cash-intensive business | `cash_intensive`, `gambling_wash`, `gambling_rapid_cycling`, `insurance_cycling` |
| Digital assets | `crypto_atm_withdrawal`, `crypto_mining_wash`, `crypto_salary_obfuscation`, `stablecoin_chain` |
| Professional & corporate abuse | `ghost_payroll`, `corporate_round_robin`, `director_loan_recycling`, `loan_back`, `hawala_mirror` |

Each typology has 7–12 scenario instances per bank. Scenario boundaries (customer IDs, account IDs, date window, signal columns) are recorded in `labels.json`.

---

## Dataset Structure

### Files

```text
oneview/
  shared/
    labels.json          # Ground-truth: all 297 unique scenarios + node labels + train/val/test splits
    dataset_card.md      # Auto-generated summary
  parquet/               # Combined Parquet files (all tables, all banks merged)
  pyg/                   # PyTorch Geometric tensors (node features, edge index, masks)
  neo4j/                 # Neo4j Cypher load files (combined graph)

<bank_id>/               # Per-bank outputs for: yellow_bank, red_bank, white_bank, orange_bank
  parquet/               # Per-bank Parquet files
  pyg/                   # Per-bank PyG tensors
  neo4j/                 # Per-bank Neo4j Cypher files
  shared/
    labels.json          # Per-bank filtered labels (cross-bank scenarios include all_account_ids)
```

### Schema (30+ tables)

Core tables across all banks:

| Table group | Tables |
|---|---|
| Customers | `customer_master`, `customer_individual`, `customer_business`, `customer_trust`, `customer_identity`, `customer_address`, `customer_kyc`, `customer_risk_profile`, `customer_financial_profile` |
| Accounts | `account_base`, `account_deposit`, `account_credit_card`, `account_loan`, `account_term_deposit`, `account_behavioral_baseline` |
| Transactions | `transaction_base`, `transaction_card_pos`, `transaction_bpay`, `transaction_transfer_internal`, `transaction_transfer_external`, `transaction_payid`, `transaction_wire_international`, `transaction_cash_deposit`, `transaction_cash_withdrawal`, `transaction_direct_debit`, `transaction_digital_context` |
| AML labels | `_aml_designations`, `_scenario_log` |

### Graph Schema

For GNN use, the dataset exports a heterogeneous graph with:

- **Node types (6):** Customer, Account, Device, IP, Beneficiary, Bank
- **Edge types (6):** Owns, Has\_Relationship, Uses, In\_Scenario, Transfers\_To, Is\_Located\_In

### `labels.json` Structure

```json
{
  "scenarios": [
    {
      "scenario_id": "SCN_0001",
      "typology": "smurfing",
      "customer_ids": ["uuid", ...],
      "account_ids":  ["uuid", ...],
      "date_from": "2026-02-16",
      "date_to":   "2026-03-06",
      "signal_columns": ["TRANSACTION_CASH_DEPOSIT.note_count", ...]
    }
  ],
  "node_labels": { "<customer_id>": 1, ... },
  "splits": { "train": [...], "val": [...], "test": [...] }
}
```

`node_labels` maps each customer UUID to `1` (AML) or `0` (legitimate). Note: the predefined train/val/test split assigns all 297 scenarios to train because the 7-month window does not provide sufficient temporal separation for held-out splits; users should define their own splits for evaluation.

---

## Customer Population

| Segment | Proportion |
|---|---|
| Individual | 75% |
| Business | 20% |
| Trust | 5% |
| Domestic (AU resident) | 85% |
| International | 15% |
| Cross-bank (2 banks) | 20% |
| Cross-bank (3 banks) | 5% |

Additional customer attributes: KYC status, risk rating, PEP/sanctions screening (2.0% PEP rate), adverse media flags (3.0%), employment type, annual income (log-normal, mean AUD 238k).

---

## Transaction Characteristics

| Metric | Value |
|---|---|
| Avg transactions / account / month | ~44 (individual deposit) |
| Weekday/weekend volume ratio | 2.6× |
| Business-hours (07:00–22:00) POS/ATM share | 80% |
| PayID proportion (individual accounts) | 8.5% |
| International wire destinations | 40 countries |
| FATF high-risk wire proportion | 15% |
| CTR-flagged cash deposits (≥ AUD 10,000) | 3.4% of cash deposits |
| Mobile transaction share | 48.8% |
| Benford's Law first-digit '1' | 33.3% (expected 30.1%) |

---


## Usage

### Load Parquet (pandas)

```python
import pandas as pd

txn  = pd.read_parquet("oneview/parquet/transaction_base.parquet")
cust = pd.read_parquet("oneview/parquet/customer_master.parquet")
```

### Load Ground-Truth Labels

```python
import json

with open("oneview/shared/labels.json") as f:
    labels = json.load(f)

# Dict mapping customer_id → 0/1
node_labels = labels["node_labels"]

# All AML scenarios with their account IDs and date windows
scenarios = labels["scenarios"]
```

### Load PyTorch Geometric Graph

```python
import torch

data = torch.load("oneview/pyg/graph.pt")
# data.x           — node feature matrix
# data.edge_index  — edge connectivity
# data.y           — node labels (0 = legitimate, 1 = AML)
# data.train_mask  — training mask
```

---

## Considerations for Using the Data

### Intended Use

- Training and benchmarking GNN-based AML detection models
- Evaluating graph anomaly detection, link prediction, and node classification algorithms in a financial-crime context
- Studying multi-bank, cross-institution typology patterns

### Out-of-Scope Use

This dataset must not be used to build systems intended to evade AML controls, launder money, or circumvent financial crime detection in production systems.

### Limitations

- **Synthetic only:** The dataset does not reflect the full complexity of real banking data, including regulatory edge cases, legacy system artefacts, or jurisdiction-specific product variations beyond the Australian configuration.
- **Balanced typology coverage:** Scenario instances per typology are capped (7–12 per bank) to ensure coverage breadth. Real AML distributions are highly skewed toward a small number of typologies.
- **Payroll–income alignment:** Synthetic payroll amounts are sampled independently from declared annual income; ~44% of customers fall within ±40% of their declared income (a known generator constraint).
- **Class imbalance:** AML customers are 5% of the population by design. Downstream models should account for this imbalance (weighted loss, oversampling, etc.).
- **No held-out test split:** All 297 scenarios fall within the training window. Users evaluating generalisation should define their own typology-held-out or time-based splits.

### Social Impact

The dataset is designed to advance automated AML detection, which supports financial crime prevention. No real individuals are represented.

---