aml-campaigngraph-alpha

AML-CampaignGraph Alpha Model Card

Model summary

CampaignGraphModel is the Alpha campaign-level model for AML-CampaignGraph. It combines an interpretable AMLGraphScorer node-evidence model with graph-level structural and temporal features, trains a class-balanced random forest, aggregates candidate campaign components, and returns a campaign score with ranked evidence transactions.

Intended use

The model is intended for research, reproducibility experiments, teaching, and comparison against future temporal graph models. It is not a production AML system.

Out-of-scope use

Do not use this model to accuse a person or institution, close an account, file a suspicious activity report, or make any legal, credit, employment, insurance, or financial decision. The model only recognizes patterns generated by the synthetic Alpha simulator.

Training data

The model is trained on AML-CampaignGraph synthetic graphs generated from the three Alpha scenarios: FAN_IN, CIRCULAR, and PASS_THROUGH, with hard negatives. The generation seed, graph-level split, and settings are versioned in the repository.

Evaluation

The primary Alpha evaluation uses independent synthetic graphs and reports node-level and campaign-level Average Precision, precision, recall, F1, abstention rate, and evidence coverage. Results must be interpreted by scenario, difficulty, and seed. The Alpha release is a baseline, not a claim of production effectiveness.

Limitations

The model does not implement the final learned temporal heterogeneous GNN proposed for the research release. It can learn simulator-specific artifacts and its inferred campaign type is heuristic. Its evidence output is a research evidence proxy rather than a causal explanation. Synthetic performance cannot establish performance on real banking data.

Responsible deployment

A responsible future deployment would require institution-specific validation, privacy review, legal and compliance review, monitoring for drift, human investigation, and a documented appeals process. None of those conditions are supplied by this Alpha release.

Citation

Cite the versioned AML-CampaignGraph repository and the research paper when available. See CITATION.cff.