Advances in Continual Graph Learning for Anti‐Money Laundering Systems: A Comprehensive Review

Bruno Deprez, Wei Wei, Wouter Verbeke, Bart Baesens, Kevin Mets, Tim Verdonck

TL;DR

This paper explores how Continual Graph Learning can prevent anti-money laundering (AML) systems from "forgetting" old fraud patterns while learning new ones. By benchmarking various strategies on synthetic and real-world datasets, the authors demonstrate that replay-based methods—specifically Gradient Episodic Memory (GEM)—are the most effective at balancing the need to adapt to new tactics without losing the ability to detect historical ones.

What Problem Does This Paper Address?

Money laundering is a $2 trillion annual problem that evolves through a constant "cat-and-mouse" game. Criminals frequently change their tactics to evade detection, requiring financial institutions to constantly fine-tune their machine learning models.

However, traditional deep learning models suffer from catastrophic forgetting: when a model is trained on new data (e.g., a new type of "layering" scheme), it often overwrites the weights used to detect previous patterns. In the AML world, this is dangerous; if a bank updates its model to catch a new Bitcoin scam, it might inadvertently stop catching classic shell-company transfers.

While Graph Neural Networks (GNNs) are excellent at tracing the "digital paper trail" of transactions, research into making them continual learners—models that accumulate knowledge over time—is still in its infancy. This paper fills that gap by evaluating how different continual learning strategies handle the extreme class imbalances and evolving patterns typical of financial crime.

Methods

The authors categorize and test three main families of continual learning strategies within a GNN framework:

  1. Replay-based (e.g., GEM): These methods maintain a small "memory buffer" of past transactions to revisit during new training sessions, ensuring the model doesn't drift too far from what it previously learned.
  2. Regularization-based (e.g., EWC, MAS, TWP): These add a penalty to the loss function that discourages the model from changing weights that were crucial for previous tasks.
  3. Architecture-based (e.g., PackNet): These isolate specific parts of the neural network for specific tasks, essentially "freezing" old knowledge in place.

The study uses a Graph Convolutional Network (GCN) as the backbone and investigates the impact of hyperparameters like model depth, width, and the number of training epochs per task.

Datasets

The study utilizes two primary benchmarks:

  • IBM AML (Synthetic): A massive dataset of over 5 million transactions. It is used for edge classification, where the goal is to identify specific laundering patterns like "fan-out" (one account to many) or "scatter-gather."
  • Elliptic (Real-world): A dataset of Bitcoin transactions grouped into 49 time intervals. It is used for node classification, identifying illicit vs. legitimate entities in a real-world blockchain environment.

Key Insights and Ideas

  • The Capacity Paradox: Deeper and wider GNNs (more layers and hidden dimensions) are generally more powerful, but they are also more prone to forgetting. The authors found that larger models tend to overfit on the most recent task, wiping out prior knowledge more aggressively than smaller, "shallower" models.
  • The "Joint" Model Failure: In many ML fields, a "Joint" model (trained on all data at once) is the gold standard. However, in AML, the authors found the Joint model often failed. Because fraud is so rare (extreme class imbalance), the Joint model simply learned to predict "not fraud" for everything to achieve high accuracy. Continual learning methods actually performed better at identifying minority fraud classes by focusing on them task-by-task.
  • Epoch Sensitivity: There is a "sudden jump" in forgetting. A model might retain old knowledge for 1–2 epochs of new training, but by epoch 5 or 10, the old knowledge is often completely overwritten.
  • Task Order Doesn't Matter (Much): Whether you show the model "easy" fraud patterns first or "hard" ones doesn't significantly change the final performance, suggesting these models are robust to the sequence of curriculum.

Key Results

  • GEM is the Winner: Replay-based methods, particularly Gradient Episodic Memory (GEM), consistently achieved the best balance of high performance (Micro-F1) and low forgetting across both datasets.
  • Negative Forgetting: In some cases on the Elliptic dataset, methods like TWP and GEM exhibited "backward transfer" (negative forgetting), meaning that learning from new Bitcoin transactions actually improved the model's ability to classify older ones.
  • Imbalance Handling: Continual learning proved to be a superior strategy for handling the 0.1% - 2% fraud prevalence compared to standard batch training.

Limitations and Future Work

  • Privacy and Regulation: Replay-based methods (the top performers) require storing a buffer of old data. In the banking sector, strict data privacy regulations (like GDPR) may limit how much historical transaction data can be stored and "replayed," potentially hindering the practical use of GEM.
  • Backbone Diversity: The study focused primarily on GCNs. The authors suggest that future work should explore more complex architectures like Graph Attention Networks (GATs), which might capture transaction nuances better but are more computationally expensive.
  • "No Fraud" Periods: Real-world systems often go through periods with zero detected fraud. Future research needs to investigate how continual learning models behave during these "dry spells" to ensure they don't lose their edge.

Summary updated

Abstract

ABSTRACT Financial institutions are required by regulation to report suspicious financial transactions related to money laundering. Therefore, they need to constantly monitor vast amounts of incoming and outgoing transactions. Given the involvement of many parties in money laundering, graph analytics is vital for effective monitoring. A particular challenge in detecting money laundering is that money launderers continuously adapt their tactics to evade detection. Hence, detection methods need constant fine‐tuning. Traditional machine learning models suffer from catastrophic forgetting when fine‐tuning the model on new data, thereby limiting their effectiveness in dynamic environments. Continual learning addresses this issue and enhances current anti‐money laundering (AML) practices by allowing models to incorporate new information while retaining prior knowledge. Research on continual graph learning for AML, however, is still scarce. In this review, we critically evaluate state‐of‐the‐art continual graph learning approaches for AML applications. We categorize methods into replay‐based, regularization‐based, and architecture‐based strategies within the graph neural network (GNN) framework, and we provide in‐depth experimental evaluations on both synthetic and real‐world AML data sets that showcase the effect of the different hyperparameters. Our analysis demonstrates that continual learning improves model adaptability and robustness in the face of extreme class imbalances and evolving fraud patterns. Finally, we outline key challenges and propose directions for future research. This article is categorized under: Data: Types and Structure > Graph and Network Data Data: Types and Structure > Streaming Data Statistical Learning and Exploratory Methods of the Data Sciences > Deep Learning
BibTeX
@inproceedings{deprez2025advances,
  author = {Bruno Deprez and Wei Wei and Wouter Verbeke and Bart Baesens and Kevin Mets and Tim Verdonck},
  title = {{Advances in Continual Graph Learning for Anti‐Money Laundering Systems: A Comprehensive Review}},
  booktitle = {Wiley},
  year = {2025},
  doi = {10.1002/wics.70040},
}

Loading related content...

Figure