AML GNN ablation

Run the four-arm GraphSAGE ablation over the transfer graph and read the result.

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

This is the hands-on runbook for pragmatiq's AML graph extension (our addition, not in the paper). See that page for why; here is how to run it.

Build a run with transfers + AML labels

pip install -e .
pragmatiq synth generate --out data/aml --n-users 12000 --n-workers 8 \
  --config configs/data/synthetic.yaml         # produces transfers.parquet + labels/aml.parquet
pragmatiq tokenize data/aml --out data/aml/tok
pragmatiq pretrain data/aml/tok --name aml --model-size small

Run the ablation

pragmatiq gnn data/aml/tok --run runs/aml \
  --transfers data/aml/transfers.parquet \
  --aml-label data/aml/labels/aml.parquet \
  --epochs 150

This trains four setups and reports held-out ROC-AUC (mean ± std over seeds): (a) a probe on isolated embeddings, (b) GraphSAGE + pragmatiq features, (c) GraphSAGE + hand-crafted features, (d) a logistic control on the hand-crafted features with no graph.

Read it

The gated claim is relational recovery — (c) 0.670 ≫ (a) 0.498: the transfer graph carries multi-hop AML signal an isolated per-user embedding cannot see, and message passing adds over the same features without a graph ((c) > (d) 0.604). The honest limitation is that the learned per-user embedding adds only a little ((b) 0.554) and does not beat hand-crafted features ((b) < (c)) — recovering the laundering signal in a learned representation is the open challenge. The full discussion is on the concept page and in the 04_aml_gnn notebook.

Build a run with transfers + AML labels

Run the ablation

Read it

On this page