Graph-feature-preprocessed gradient boosting
GFP (graph features) + XGBoost
Part of QSMPC-QKD-QHE-AI-Hybrid,
a quantum-safe orchestration demo.
This is the plaintext model for the finance_aml use case; the encrypted
path runs a distilled student, not this model.
Measured performance
| metric | value |
|---|---|
accuracy | 0.9982855312490154 |
auprc | 0.12334101789973353 |
auprc_baseline | 0.0015234617040716665 |
auroc | 0.9525582593503528 |
decision_threshold | 0.386869 |
minority_class_f1 | 0.15194805194805194 |
n_test | 1523504 |
n_train | 3554841 |
positive_rate | 0.00101943 |
wall_clock_s | 96.6 |
Published baselines this is measured against
- Target metric: minority-class F1 (laundering)
- Baseline to beat: 0.6323 - GFP+XGBoost, IBM AMLworld HI-Small, NeurIPS 2023 D&B (63.23 +/- 0.17)
- Published ceiling: 0.7401 - MEGA-PNA, arXiv:2412.00241v2 Table 2 'Minority-class F1 scores (%) on AML edge classification task', AML Small HI (74.01 +/- 1.55). PREPRINT - no venue confirmed, which is why it is named as such. It replaces Multi-PNA+EU (AAAI-24, 68.16 +/- 2.65): that row is in the SAME table with two better ones above it (MEGA-GIN 70.83, MEGA-PNA 74.01) and has not been the ceiling since December 2024. Verified by reading the table, not the abstract.
- Companion metric shown alongside:
auprc- reported together because the aggregate figure can look healthy while the class that matters is not.
Gap to the published baseline
This reaches minority-class F1 well below the published 0.6323 (GFP+XGBoost, AMLworld HI-Small, NeurIPS 2023 D&B). The cause is a feature ARCHITECTURE difference, not a tuning one: the gap is unreachable at any threshold, because measured AUPRC is far below what that F1 requires. IBM's Graph Feature Preprocessor computes per-transaction subgraph patterns over a dynamically updated graph; this project computes per-account scalar aggregates. A degree is one number per account, whereas a GFP feature is a pattern match in that transaction's own neighbourhood at the moment it arrives. GFP ships in snapml under a Proprietary licence and is excluded by this project's commercial-OK-only rule, but the licence is not itself the cause - the architecture is, and it is reimplementable without snapml, simply at more cost than renaming aggregates. Reachability is computed to 4 hops against injected CYCLE attempts of up to 10; that was previously recorded as a main cause and is NOT one - the GFP paper (arXiv:2402.08593) decomposes its own contribution and caps multi-hop cycle and scatter-gather features at about 4% relative, while fan-in/fan-out, which is reimplemented here, carries more than 30%. The evaluation is strictly inductive and temporally split. Per-account aggregates are causal (each row sees only edges strictly before it); the structural block - cycle closure, scatter-gather, bank diversity, burst degree - is still fitted on the training window, because recomputing sparse reachability at every timestep is not tractable here. Reported as measured; not adjusted to close the gap. Full workings: references/finance-gap-investigation.md.
Training data
- Dataset: IBM AMLworld HI-Small
- Licence: CDLA-Sharing-1.0
- Source: https://www.kaggle.com/datasets/ealtman2019/ibm-transactions-for-anti-money-laundering-aml (licence read 2026-08-03)
- Attribution (CDLA-Sharing-1.0): Altman et al., 'Realistic Synthetic Financial Transactions for Anti-Money Laundering Models', NeurIPS 2023 Datasets and Benchmarks. Data: IBM Transactions for Anti-Money Laundering (AML), CDLA-Sharing-1.0.
5M transactions, 515K accounts, ~1 laundering transaction in 981. NeurIPS 2023 D&B. SYNTHETIC: IBM generates it with a multi-agent virtual-world model and states 'Everything is synthetic'; the paper is titled 'Realistic Synthetic Financial Transactions for AML Models'. Chosen for its calibration to real transaction statistics and its CDLA-Sharing terms, not because it is real data - no commercial-OK corpus of real laundering transactions exists.
Notes and limitations
Stateless at inference, so this is the browser-tier finance model.
Honest scope
This model is published as part of a research proof of concept, not as a production system.
Numbers above are what this repository measured on the split described, with the code in
scripts/train/. Where a figure is carried from the literature rather than measured here,
it is labelled as such.