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. 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 methods may address this issue and enhance 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 categorise 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.
Introduction
Money laundering is a global problem involving approximately $2 trillion annually. To combat this, financial institutions utilize machine learning to monitor transactions. However, criminals constantly evolve their tactics, creating a "cat-and-mouse" game that requires models to be frequently updated.
Standard deep learning models suffer from catastrophic forgetting: when trained on new data (e.g., a new Bitcoin scam), they often overwrite the weights used to detect previous patterns (e.g., classic shell-company transfers). While Graph Neural Networks (GNNs) are ideal for tracing transaction trails, making them continual learners that accumulate knowledge over time is a relatively new field of research. This paper evaluates how different continual learning strategies handle the extreme class imbalances and evolving patterns typical of financial crime.
Continual Learning Strategies
The study evaluates three main families of continual learning strategies within a GNN framework:
- Replay-based (e.g., GEM): These methods maintain a small "memory buffer" of past transactions to revisit during new training sessions, ensuring the model retains previous knowledge.
- Regularization-based (e.g., EWC, MAS, TWP): These add a penalty to the loss function to discourage the model from changing weights that were crucial for previous tasks.
- Architecture-based (e.g., PackNet): These isolate specific parts of the neural network for specific tasks, effectively "freezing" old knowledge.
Experimental Insights
The authors benchmarked these strategies using a Graph Convolutional Network (GCN) backbone on two primary datasets: the synthetic IBM AML dataset (5 million transactions) and the real-world Elliptic Bitcoin dataset. Several key insights emerged:
- The Capacity Paradox: Larger, deeper GNNs are more powerful but also more prone to forgetting. They tend to overfit on the most recent task, wiping out prior knowledge more aggressively than smaller models.
- Superiority Over Joint Models: In many ML fields, a "Joint" model trained on all data at once is the gold standard. However, in AML, where fraud is extremely rare (0.1% - 2%), Joint models often fail by simply learning to predict "not fraud" for everything. Continual learning methods performed better by focusing on minority fraud classes task-by-task.
- Epoch Sensitivity: Forgetting often happens suddenly. A model might retain old knowledge for 1–2 epochs of new training, but by epoch 10, the old knowledge is frequently completely overwritten.
Key Results and Future Work
Gradient Episodic Memory (GEM), a replay-based method, consistently achieved the best balance of high performance and low forgetting. Interestingly, on the Elliptic dataset, some methods exhibited "backward transfer," where learning from new transactions actually improved the model's ability to classify older ones.
However, practical implementation faces challenges. Replay-based methods require storing historical data, which may conflict with privacy regulations like GDPR. Future research must address these privacy concerns and investigate how models behave during "dry spells" where no fraud is detected for extended periods.
Summary updated