AML over the transfer graph

pragmatiq is an independent implementation inspired by the PRAGMA paper (arXiv 2604.08649) and is not affiliated with or endorsed by Revolut.

Why a graph

Mule-ring (money-laundering) detection is relational: a launderer is defined by who they transact with — fan-in of small credits, layering among the ring, rapid fan-out — not just their own behavior. A per-user embedding, however good, cannot by construction see a counterparty pattern that spans accounts. So we build a directed transfer graph from transfers.parquet and run a four-arm comparison.

The four-arm ablation

api.gnn(...) trains a 2–3 layer GraphSAGE and reports held-out ROC-AUC, mean ± std over seeds, for (full-scale benchmark: 12k accounts × 3 seeds, small model):

• (a) isolated pragmatiq embeddings — a probe on each user's embedding alone, no graph. 0.498
• (b) GraphSAGE + pragmatiq features — message passing over the transfer graph with pragmatiq embeddings as node features. 0.554
• (c) GraphSAGE + hand-crafted features — the same graph, with generic structural node features (degree, volume, counterparties). 0.670
• (d) logistic control — the same hand-crafted features, no graph. 0.604

The synthetic mule rings are multi-hop layered laundering chains: mule amounts and counterparty degree are drawn to match ordinary accounts, so 1-hop degree is not a trivial oracle (the degree-only control (d) reaches only 0.604). The gated result is relational recovery: a GraphSAGE over the transfer graph recovers rings that a probe on the isolated embedding cannot — (c) 0.670 ≫ (a) 0.498 — so the AML signal lives in the multi-hop transfer structure an isolated embedding misses, and message passing adds over the same features without a graph ((c) > (d)).

pip install -e .
pragmatiq gnn data/tok --run runs/aml \
  --transfers data/synth/transfers.parquet \
  --aml-label data/synth/labels/aml.parquet

The honest limitation

The learned per-user embedding adds only a little over the isolated probe ((b) 0.554 > (a) 0.498) and does not beat hand-crafted features ((b) 0.554 < (c) 0.670). The isolated embedding sits near chance, so the model does not capture the multi-hop laundering signal in the per-user representation; recovering it in a learned representation is the open challenge. This is consistent with the PRAGMA paper's own observation that AML is a setting where the model underperforms because it processes user histories in isolation — the GNN here is pragmatiq's honest extension that probes that gap, not a "the learned embedding wins" result. See MODEL_CARD.md and the 04_aml_gnn notebook for the full discussion and the latest numbers.