# Notices

This dataset is derived data under the **Community Data License Agreement,
Sharing, Version 1.0**. The agreement text is in
[`LICENSE-CDLA-SHARING-1.0.txt`](LICENSE-CDLA-SHARING-1.0.txt) and at
<https://cdla.dev/sharing-1-0/>. Section 3.3 asks that anyone publishing data
under it include the text, the name, or a link; all three are here.

If you publish anything derived from this dataset, the same terms bind you. That
is what "Sharing" means in the name, and it is the practical difference between
this and a permissive licence.

## Source

Every derived artefact here comes from **AMLworld**, IBM's synthetic
anti-money-laundering transaction dataset (Altman et al., *Realistic Synthetic
Financial Transactions for Anti-Money Laundering Models*, NeurIPS 2023 Datasets
and Benchmarks), published under CDLA-Sharing-1.0 at
<https://www.kaggle.com/datasets/ealtman2019/ibm-transactions-for-anti-money-laundering-aml>.
Licence confirmed from that listing on 2026-08-10.

The original AMLworld transaction CSVs are **not** redistributed here. Fetch
them from the listing above under your own Kaggle account.

## Changes made to the Data

Section 3.1(b) asks for prominent notice of changed files. Nothing here is a
copy of an AMLworld file; everything is computed from one. In the order the
pipeline applies them:

1. **Selection.** The final 20% of HI-Small and LI-Small rows under the source
   CSV file-order 60/20/20 partition was reduced to 3,753 and 2,268 edges by
   HT-Coreset with sampling seed 0: every illicit edge retained, benign edges
   sampled by supervised difficulty. The 1:2 illicit-to-benign ratio uses the
   smallest benign budget evaluated in the construction study. Timestamps
   were not sorted, and time ranges overlap across the partitions.
   `extras/*/ht_subset_indices.npy` records which rows survived.
2. **Weighting.** Each retained edge carries an inverse-inclusion-probability
   weight in `extras/*/ht_weights.npy`. The weights sum to the full split size,
   1,015,669 and 1,384,810. These weights were computed for the selected
   evaluation cases; they are not original AMLworld fields.
3. **Serialisation.** Graph extraction started at each target edge's source
   account and expanded over two hops, capped at 50 incoming/outgoing
   neighbours together per expanded account per hop, then flattened the graph
   into `typed_graph_text`. That column contains re-expressed AMLworld
   transaction values. The target was not explicitly retained or marked; see
   the [archived input defects](README.md#archived-input-defects).
4. **Feature extraction.** `extras/*/features_gfp.npz` holds 79 columns per
   edge produced by the Snap ML Graph Feature Preprocessor over the AMLworld
   graph.
5. **Relabelling of identifiers.** Cases were retagged from `v2_NNNNN` to
   `amlc_NNNNN` on 2026-08-10. The digits are unchanged and are the row
   position. Recorded model outputs still quote the earlier form, because they
   are recorded output and were not rewritten.
6. **Corrections after the first build.** LI-Small HT weights were corrected,
   and typology decoding was aligned with the released integer-to-label mapping.
7. **Evaluation scorer update (2026-09-14).** `ensemble_probs_coreset.npy`
   now averages LightGBM+GFP and XGBoost+GFP models trained on the file-order
   partition.
   The original three-model construction scores are preserved in
   `construction_probs_coreset.npy`; they include GCPAL+GFP trained with a
   random split that overlaps the file-order test partition. `scoring_metadata.json`
   records both roles. Targets, graph text, labels, and inclusion weights
   are unchanged by this update.
8. **Documentation update (2026-09-30).** The README, this notice, and scoring
   metadata incorporate the 2026-09-28 audit findings: the CSV file-order
   partition, sampling seed, and target-edge and time-summary defects in the
   archived graph strings. The README distinguishes HT-weighted full-test
   estimates from unweighted subset diagnostics computed from the same
   predictions. It also clarifies the selected budget, inherited test-selected
   thresholds, and archived labels used for published TF1 values. Graph text,
   labels, weights, features, indices, and model probabilities are unchanged.

## What is not CDLA

Section 3.5 says the agreement "imposes no obligations or restrictions on Your
Use or Publication of Results", and §1.11 limits Results to no more than a de
minimis portion of the Data. On that basis:

- **`ml_baselines/weights/`** and **`ml_baselines/test_probs/`** are model
  parameters and model outputs, not Data. They are ours and are released under
  the **MIT** licence, the same as the code.
- The aggregate metric tables in the code repository's `results/` are Results.

Everything under `data/` and `extras/` is Data and is CDLA-Sharing-1.0.

## Elliptic

No Elliptic artefact ships here. The Elliptic Bitcoin dataset is CC BY-NC-ND
4.0, whose NoDerivatives clause forbids redistributing adapted material, so the
code repository ships the pipeline that reproduces those experiments from the
original release and nothing derived from it.