adnanallemon/aml-campaigngraph-data

AML-CampaignGraph Data Card

Dataset summary

AML-CampaignGraph v1.0.0 is a fully synthetic temporal transaction-graph benchmark for campaign-level anti-money-laundering research. It is designed for reproducible experiments on campaign ranking, temporal generalization, controlled out-of-distribution evaluation, and evidence extraction. The benchmark deliberately uses generated data because public banking records contain sensitive financial and personal information 1.

The dataset contains 180 graph instances: 90 positive campaign graphs and 90 hard-negative graphs. Each of the six scenarios contributes 15 positive and 15 hard-negative graphs. The release contains 4,320 synthetic nodes and 14,790 timestamped transactions.

Synthetic-data notice: No row represents a real customer, account, bank, institution, transaction, or confirmed criminal case. A label is a generator-defined pattern label, not a legal or investigative finding.

Motivation and research use

Most AML resources operate at the alert or transaction level. AML-CampaignGraph provides a controlled campaign-level unit in which accounts and transactions form a directed temporal graph. It is intended for reproducibility research, temporal graph feature engineering, interpretable ML, fraud and financial-crime education, benchmark construction, and explanation evaluation.

The benchmark is complementary to synthetic AML resources such as SynthAML 1 and IBM AMLSim 2. It does not claim to replace real institution-specific validation or to reproduce the full complexity of regulated financial monitoring.

Dataset composition

Property	Value
Version	1.0.0
Format	aml-campaigngraph-final-csv
Graphs	180
Positive campaign graphs	90
Hard-negative graphs	90
Nodes	4,320
Transactions	14,790
Scenarios	6
Nodes per graph	24
Temporal horizon	60 days
Temporal windows	6
Generation seed for published bundle	2025
Difficulty	medium

Scenario	Positive graphs	Hard-negative graphs	Total
FAN_IN	15	15	30
FAN_OUT	15	15	30
CIRCULAR	15	15	30
PASS_THROUGH	15	15	30
LAYERED	15	15	30
SMURFING	15	15	30
Total	90	90	180

Campaign patterns

The generator implements six named patterns that represent controlled graph motifs rather than real-world typologies. FAN_IN concentrates incoming activity, FAN_OUT concentrates outgoing activity, CIRCULAR creates a directed loop, PASS_THROUGH connects upstream and downstream activity through an intermediary, LAYERED creates multiple sequential layers, and SMURFING creates multiple low-value fragments that converge into a campaign structure. The scenario name is metadata supplied by the generator and should not be interpreted as a regulatory classification.

Each positive graph has campaign membership and campaign-edge labels. Hard negatives are generated to have comparable graph size and activity volume while omitting the labelled campaign motif. Difficulty controls the separation between campaign and background behavior.

Files and schema

File	Key fields	Description
nodes.csv	graph_id, node_id, node_type, region_bucket, is_business, is_campaign_node, campaign_role	Synthetic account/entity nodes and node-level campaign labels.
transactions.csv	graph_id, transaction_id, src_id, dst_id, timestamp, amount, channel, currency_bucket, status, is_campaign_edge, motif_label, motif_role, time_window	Directed, timestamped synthetic transaction edges and motif metadata.
campaigns.csv	graph_id, campaign_id, campaign_type, is_campaign, is_hard_negative, start_time, end_time, severity, generation_seed, difficulty	One graph-level campaign record and its label.
campaign_membership.csv	graph_id, campaign_id, node_id, role, is_core	Campaign-node membership and role labels.
alerts.csv	graph_id, alert_id, trigger_time, alert_rule, review_priority, campaign_id	Synthetic monitoring-alert metadata for research workflows.
metadata.json	format, version, counts, scenarios	Machine-readable release summary and provenance.

Identifiers are synthetic and scoped to generated graphs. Amounts, regions, currencies, channels, statuses, and alert rules are simulator fields or buckets; they are not calibrated to a particular jurisdiction or institution.

Generation and provenance

The generator is implemented in src/aml_campaigngraph/generator.py. The complete source code, experiment scripts, paper manuscript, and versioned release are available in the GitHub repository. The published data bundle can be regenerated with:

pip install -e ".[dev]"
python scripts/build_final_dataset.py

The exact final settings are stored in configs/final.yaml. The published bundle uses explicit integer seeds, six scenarios, 24 nodes per graph, a 60-day horizon, six temporal windows, medium difficulty, and a 0.5 hard-negative ratio. The split manifests in artifacts/final/split_manifest.json are generated by complete graph_id and are not row-level random splits.

Evaluation protocols

The benchmark supports three graph-disjoint protocols. primary is a stratified holdout. temporal holds out a future time window. ood holds out a scenario-and-difficulty combination to measure controlled distribution shift. No node or transaction row from a test graph should appear in training. Researchers should report results by protocol, scenario, difficulty, and seed rather than only reporting a pooled score.

Intended use

The dataset is suitable for algorithm development, temporal feature engineering, campaign-level ranking, interpretable baseline comparison, calibration studies, selective prediction, graph-disjoint evaluation, evidence overlap, and teaching. It can also be used to test whether a method fails gracefully under hard negatives and controlled distribution shift.

Out-of-scope use

Do not use this dataset to identify real people, infer criminality, make account or customer decisions, file regulatory reports, create legal conclusions, or claim that a model trained on it represents a real bank. Do not join the synthetic identifiers with external personal data. Do not present a high benchmark score as evidence of operational AML effectiveness.

Biases, limitations, and known risks

The simulator encodes design choices that may make the task easier or harder than real monitoring. The patterns are compact, the label semantics are generator-defined, and the distributions of amounts, channels, regions, and timing are synthetic. Legitimate operational behavior may be underrepresented, and real laundering behavior is more heterogeneous than the six controlled patterns. The benchmark can therefore measure reproducibility and method sensitivity, but not institutional risk or regulatory performance.

The final experiments show that several models obtain very high ranking performance on the controlled data, while OOD F1 and calibration are weaker. This is an important limitation rather than a deployment recommendation. Researchers should include negative controls, alternative generators, real-world validation where lawfully available, and uncertainty analysis in subsequent work.

Privacy and safety

The dataset contains no personal data and no real financial records. The associated Space displays the same warning and should be used only with synthetic inputs. If a researcher adapts the generator with external data, that adaptation must be independently reviewed for privacy, licensing, consent, and security before release.

License

The code and generated benchmark are released under the MIT license. See LICENSE for the complete terms.

Citation

Please cite AML-CampaignGraph v1.0.0 using CITATION.cff, and cite the accompanying research manuscript when using the benchmark, model, or evaluation protocol.

References