Networked Markets, Fragmented Data: Adaptive Graph Learning for Customer Risk Analytics and Policy Design

Lecheng Zheng, Jian Ni, Chris Zobel, John R. BirgearXiv 2025 · DOI · arXiv

Introduction

Anti-Money Laundering (AML) is a critical challenge for global financial stability, with illicit flows estimated between $800 billion and $2 trillion annually. Traditional detection methods are increasingly outmatched by sophisticated criminal networks that exploit the fragmented nature of the global banking system. This paper introduces an integrated framework that leverages Federated Learning (FL), Graph Neural Networks (GNNs), and Hierarchical Reinforcement Learning (HRL) to detect cross-bank money laundering while preserving data privacy. By combining these technologies, the authors provide a blueprint for how competing financial institutions can collaborate to catch sophisticated criminals without ever sharing sensitive customer data.

The Challenge of Fragmented Data

Money laundering is inherently a networked crime. Criminals use tactics such as "structuring" or "smurfing" to break large sums into smaller transactions across multiple institutions and jurisdictions. This creates three primary obstacles for modern financial institutions:

  1. Data Silos: Privacy regulations and competitive concerns prevent banks from sharing raw transaction data. Criminals exploit this by fragmenting their activities so that no single institution sees the full picture.
  2. Extreme Imbalance: Suspicious transactions account for less than 1% of total volume. Standard AI models often struggle to identify these rare events without generating an overwhelming number of false positives.
  3. Static Policies: Most banks rely on rigid, rule-based systems (e.g., flagging any transfer over $10,000). These fail to adapt to evolving criminal tactics and often cause unnecessary friction for legitimate customers.

Methodology: An Integrated Framework

The researchers propose a multi-layered "Customer Intelligence" framework consisting of three main modules designed to address these hurdles.

Federated Graph Neural Network (FedGNN)

To overcome data silos, the system uses a Federated Learning approach where each bank trains a GNN on its local transaction graph. To capture cross-bank relationships without sharing raw data, the authors introduce virtual supernodes. These act as boundary connectors that share abstracted, privacy-preserving embeddings of cross-border activities. Additionally, the model utilizes Focal Loss to force the neural network to prioritize learning from rare, hard-to-detect suspicious cases rather than being overwhelmed by the majority of legitimate transactions.

Cross-bank Personalized PageRank (PPR)

To ensure the model's output is actionable for human risk managers, the system employs a modified PageRank algorithm. This spreads "suspiciousness scores" through the network to identify coordinated clusters. This method reveals classic laundering structures such as "fan-out" (one-to-many), "loops," and "gather-scatter" (many-to-one-to-many) patterns, providing much-needed interpretability to the AI's findings.

Hierarchical Reinforcement Learning (HRL)

Instead of relying on a fixed threshold for flagging transactions, an RL agent learns when to monitor, escalate, or freeze an account. A global coordinator ensures that local bank policies stay aligned with the broader network's goals, balancing the cost of fraud against the operational friction caused to legitimate customers.

Experimental Setup and Datasets

The study utilizes the IBM Anti-Money Laundering (AML) Small dataset, a high-fidelity synthetic dataset that mimics real-world financial behavior and includes injected laundering schemes.

  • Scale: The primary analysis was conducted on a subset of 1.4 million transactions across seven markets (US, Germany, France, Italy, Spain, China, and "Rest Countries").
  • Scalability: The authors validated the detection module on a larger 5 million transaction dataset to ensure the framework can handle enterprise-level volumes.

Key Insights

  • The Importance of Global Scaling: A significant finding was that standard local data normalization (e.g., normalizing transaction amounts within a single country) actually degrades performance by hiding global patterns. The authors recommend a "fixed-value min-max" strategy to maintain a consistent scale across the federated network.
  • Structure Over Features: The FedGNN consistently outperformed feature-only models. This confirms that how accounts interact is more predictive of criminal activity than the individual attributes of the accounts themselves.
  • Communication Efficiency: In federated learning, syncing the global model too frequently is counterproductive. Moderate communication intervals provided the best balance between accuracy and synchronization costs.

Results

The proposed framework demonstrated significant improvements over traditional, siloed models:

  • Accuracy: The framework reduced false positive and false negative rates to 6.02% and 5.53%, respectively.
  • Loss Prevention: The HRL module prevented 79.25% of potential losses, compared to only 49.41% under traditional fixed-rule policies.
  • Collaborative Gain: Sharing intelligence across markets increased the precision of identifying high-risk customer networks by 60%.

Conclusion

By combining privacy-preserving collaboration with adaptive learning, this research provides a scalable blueprint for modern financial surveillance. While the model currently relies on some labeled data, future work aims to adapt the framework for real-time streaming environments and incorporate multi-modal data, such as customer profiles and text-based suspicious activity reports, to further sharpen detection capabilities.

Summary updated

Abstract

Financial institutions face escalating challenges in identifying high-risk customer behaviors within massive transaction networks, where fraudulent activities exploit market fragmentation and institutional boundaries. We address three fundamental problems in customer risk analytics: data silos preventing holistic relationship assessment, extreme behavioral class imbalance, and suboptimal customer intervention strategies that fail to balance compliance costs with relationship value. We develop an integrated customer intelligence framework combining federated learning, relational network analysis, and adaptive targeting policies. Our federated graph neural network enables collaborative behavior modeling across competing institutions without compromising proprietary customer data, using privacy-preserving embeddings to capture cross-market relational patterns. We introduce cross-bank Personalized PageRank to identify coordinated behavioral clusters providing interpretable customer network segmentation for risk managers. A hierarchical reinforcement learning mechanism optimizes dynamic intervention targeting, calibrating escalation policies to maximize prevention value while minimizing customer friction and operational costs. Analyzing 1.4 million customer transactions across seven markets, our approach reduces false positive and false negative rates to 4.64% and 11.07%, substantially outperforming single-institution models. The framework prevents 79.25% of potential losses versus 49.41% under fixed-rule policies, with optimal market-specific targeting thresholds reflecting heterogeneous customer base characteristics. These findings demonstrate that federated customer analytics materially improve both risk management effectiveness and customer relationship outcomes in networked competitive markets.
BibTeX
@misc{zheng2025networked,
  author = {Lecheng Zheng and Jian Ni and Chris Zobel and John R. Birge},
  title = {{Networked Markets, Fragmented Data: Adaptive Graph Learning for Customer Risk Analytics and Policy Design}},
  year = {2025},
  eprint = {2512.24487},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.24487},
}
Figure