Money-laundering detection on 63M+ synthetic financial transactions, built as a 9-stage ML pipeline: data cleaning, EDA, graph feature engineering, model selection, training, evaluation, and a bias-mitigation ablation. Full stage-by-stage writeups, scripts, and rationale live in ML_Cycle/, starting at 1-Problem_Definition.
Model selection (stage 6) compares candidates on a fast, subsampled dataset (3.23% positive) for speed. That number is optimistic: the true laundering rate is roughly 1 in 1,200 transactions (0.082%). Stage 7 re-measures the selected model on 6.3M held-out transactions at the real rate to get production-honest numbers.
Comparing a plain baseline against scale_pos_weight-weighted configs on top
of already-subsampled data (stage 6), the weighted configs look fine on
PR-AUC but collapse on F1 at a fixed threshold. Negative subsampling and
scale_pos_weight are both imbalance corrections; stacking them
over-corrects.
The trained model's true positives are 99.9% Payment Format == ACH. Stage 9
tests whether that's a shortcut by retraining with the column removed
entirely. The model still catches almost exclusively ACH transactions, even
without ever seeing the label.
Removing the column costs overall PR-AUC and F1 but improves recall at a strict 90%-precision operating point. That points to the ACH concentration being a genuine property of how the underlying laundering patterns route through the data, not a shortcut a model could be trivially engineered out of.
| Stage | Metric | Value |
|---|---|---|
| Stage 6: Model Selection (subsampled) | PR-AUC / F1 @ 0.5 / Recall @ 90% precision | 0.840 / 0.785 / 0.585 |
| Stage 7: Model Training (real-world eval) | PR-AUC / F1 @ 0.5 / Recall @ 90% precision | 0.419 / 0.308 / 0.122 |
| Stage 8: Model Evaluation | Best-F1 threshold (vs. arbitrary 0.5) | 0.822 (Precision 0.618, Recall 0.364, F1 0.458) |
| Stage 8: Model Evaluation | Inference latency (p50 / p95) | 2.70 ms / 8.90 ms |
| Stage 8: Model Evaluation | Batch throughput @ 100K rows | 2.02M transactions/sec |
| Stage 9: Bias Mitigation (ablated, no Payment Format) | PR-AUC / F1 @ 0.5 / Recall @ 90% precision | 0.360 / 0.280 / 0.164 |
Full methodology, rationale, and honest caveats for every number above are in each stage's own README under ML_Cycle/.
Source: IBM Transactions for Anti Money Laundering (AML) (Kaggle).
Downloaded subset: Medium, both illicit-ratio variants (HI = higher illicit ratio, LI = lower illicit ratio), stored in dataset/.
Transaction records.
| Column | Description |
|---|---|
Timestamp |
Date and time of the transaction |
From Bank |
Bank ID of the sending account |
Account (1st) |
Account number of the sender |
To Bank |
Bank ID of the receiving account |
Account (2nd) |
Account number of the receiver |
Amount Received |
Transaction amount as received (in receiving currency) |
Receiving Currency |
Currency of the amount received |
Amount Paid |
Transaction amount as paid (in payment currency) |
Payment Currency |
Currency of the amount paid |
Payment Format |
Payment method/channel (e.g. ACH, Wire, Credit Card, Cheque) |
Is Laundering |
Label: 1 if the transaction is part of a laundering pattern, else 0 |
Note: the raw CSV header repeats the column name Account for both the sender and receiver account fields.
Account-to-entity mapping.
| Column | Description |
|---|---|
Bank Name |
Name of the bank holding the account |
Bank ID |
Numeric ID of the bank (matches From Bank / To Bank in the transactions file) |
Account Number |
Account number (matches Account in the transactions file) |
Entity ID |
ID of the entity (individual/organization) that owns the account |
Entity Name |
Name of the owning entity |
Plain-text listing of the labeled laundering patterns embedded in the transaction data. Each block is delimited by BEGIN LAUNDERING ATTEMPT - <PATTERN TYPE> / END LAUNDERING ATTEMPT - <PATTERN TYPE> and contains the constituent transactions in the same column order as the Trans.csv files:
Timestamp, From Bank, From Account, To Bank, To Account, Amount Received, Receiving Currency, Amount Paid, Payment Currency, Payment Format, Is Laundering
Pattern types observed include STACK, CYCLE, and FAN-IN (fan-out and other topologies also appear in the full dataset).