References and related work#

Level: Expert

You will: find the standards, libraries, papers, and projects that inform

FraudTwin’s design without treating them as runtime dependencies. Before you start: the Architecture guide.

Services: None.

FraudTwin is released under the Apache License 2.0.

These references informed FraudTwin’s design. The repository does not bundle copied proprietary source code or external datasets; generated records are synthetic and produced by FraudTwin. Third-party dependencies and referenced works remain subject to their own licenses and terms.

Mothilal, Sharma & Tan, Explaining Machine Learning Classifiers through Diverse Counterfactual Explanations (design reference only) - informed proximity, feasibility, and constrained counterfactual search; no source code, data, or models are copied.
DiCE reference implementation (design reference only) - informed pluggable distance and feasibility concepts; FraudTwin does not import or depend on the repository.

Synthetic data and fraud workflows#

Fraud Detection Handbook simulator (handbook) - informed deterministic customer profiles, temporal payment behavior, rule-based fraud scenarios, and the legitimate baseline used for hard negatives.
Fraud Detection Handbook validation strategies (handbook) - informed point-in-time features, delayed-label gaps, future-only evaluation windows, and rolling backtests.
Vasquez et al., The Hidden Cost of Fraud (paper) - motivated positive-unlabeled handling for fraud that remains undetected.
He et al., Identifying Labeling Mechanism in Positive-Unlabeled Learning under Unknown Class Prior (paper) - informed selection-dependent labeling controls.
PaySim (repository and simulator) - provided comparison points for synthetic mobile-money behavior and aggregate calibration without importing its schemas or data.
SynthFin Core (repository) - provided practical comparison points for behavioral enrichment, fraud scenarios, and fraud-pattern generation.

Payment and financial references#

TigerBeetle financial accounting and two-phase transfers (technical references) - informed double-entry ledger invariants and authorization-hold, post, and void semantics.
Stripe manual capture, PaymentIntent lifecycle, and disputes (public API documentation) - informed the card authorization, capture, reversal, refund, and chargeback lifecycle abstractions.
Banco Central do Brasil Pix initiation standards, Pix/SPI technical documents, and the Pix timing manual (official standards) - informed the Pix-like lifecycle and timing model without claiming to reproduce proprietary SPI internals.

Graph fraud and research papers#

Santander Gen-Fraud-Graph (repository) - informed reproducible financial graph generation, fraud-ring patterns, graph exports, and benchmark-oriented scale considerations.
NumPy parallel random generation and SeedSequence (technical references) - informed hierarchical deterministic streams for independent scale workers.
IBM AMLSim (repository) - provided a comparison point for multi-agent synthetic banking graphs and known AML/fraud patterns.
Haghighi et al., Beyond pairwise relationships: a transformer-based hypergraph learning approach for fraud detection (paper) - motivated higher-order graph relations and the separation of observable structure from latent fraud truth.
Prasetya et al., A multi-rounded adversarial scenario for graph-based promo fraud detection (paper) - informed evolving, multi-round difficulty and adversarial graph stress controls.
Fan et al., Fraud learns too (paper) - informed strategic drift and changing fraud-network structure as future-facing stress dimensions.
GRAD: Guided Relation Diffusion Generation for Graph Augmentation in Graph Fraud Detection (paper) - informed relation-level augmentation and benign-looking graph support events.
Fraud detection based on GNNs with local augmentation and adaptive relation aggregation (paper) - informed the distinction between feature camouflage and relation camouflage controls.
FRAUDAR: Graph-based fraud detection in the face of camouflage (paper) - provided the foundational relation-camouflage threat model for making fraudulent structure resemble legitimate activity.

Engineering references#

scikit-learn model evaluation - informed ranking, threshold, calibration, and classification metrics.
MLflow Tracking - informed optional experiment lineage and model-artifact logging.

Reference calibration#

SDV data quality reports and SDMetrics quality reports - informed aggregate distribution, pair-trend, cardinality, and fidelity-report concepts. FraudTwin stores deterministic summaries rather than source rows.
PaySim - informed aggregate calibration of simulation parameters while retaining domain-specific causal payment rules.

Advanced Campaign Dynamics references#

Prasetya et al., A multi-rounded adversarial scenario for graph-based promo fraud detection - evolving campaign structure.
Fan et al., Fraud learns too - structural drift and perturbation concepts.
Haghighi et al., Beyond pairwise relationships - higher-order fraud relationships.
Python Packaging User Guide, Pydantic documentation, and pytest documentation - informed package layout, typed configuration validation, and deterministic regression testing.

Next#

Return to Architecture to connect these references to the design, or use Integrate and serve for implementation routes.