Academic

Normalisation and Initialisation Strategies for Graph Neural Networks in Blockchain Anomaly Detection

arXiv:2602.23599v1 Announce Type: new Abstract: Graph neural networks (GNNs) offer a principled approach to financial fraud detection by jointly learning from node features and transaction graph topology. However, their effectiveness on real-world anti-money laundering (AML) benchmarks depends critically on training practices such as specifically weight initialisation and normalisation that remain underexplored. We present a systematic ablation of initialisation and normalisation strategies across three GNN architectures (GCN, GAT, and GraphSAGE) on the Elliptic Bitcoin dataset. Our experiments reveal that initialisation and normalisation are architecture-dependent: GraphSAGE achieves the strongest performance with Xavier initialisation alone, GAT benefits most from combining GraphNorm with Xavier initialisation, while GCN shows limited sensitivity to these modifications. These findings offer practical, architecture-specific guidance for deploying GNNs in AML pipelines for datasets wi

D
Dang Sy Duy, Nguyen Duy Chien, Kapil Dev, Jeff Nijsse · · 1 min read · 11 views
#cs.LG

arXiv:2602.23599v1 Announce Type: new Abstract: Graph neural networks (GNNs) offer a principled approach to financial fraud detection by jointly learning from node features and transaction graph topology. However, their effectiveness on real-world anti-money laundering (AML) benchmarks depends critically on training practices such as specifically weight initialisation and normalisation that remain underexplored. We present a systematic ablation of initialisation and normalisation strategies across three GNN architectures (GCN, GAT, and GraphSAGE) on the Elliptic Bitcoin dataset. Our experiments reveal that initialisation and normalisation are architecture-dependent: GraphSAGE achieves the strongest performance with Xavier initialisation alone, GAT benefits most from combining GraphNorm with Xavier initialisation, while GCN shows limited sensitivity to these modifications. These findings offer practical, architecture-specific guidance for deploying GNNs in AML pipelines for datasets with severe class imbalance. We release a reproducible experimental framework with temporal data splits, seeded runs, and full ablation results.

Executive Summary

This article presents a systematic ablation study of initialisation and normalisation strategies for graph neural networks (GNNs) in blockchain anomaly detection, specifically for anti-money laundering (AML) benchmarks. The authors investigate the effectiveness of three GNN architectures (GCN, GAT, and GraphSAGE) with various weight initialisation and normalisation techniques on the Elliptic Bitcoin dataset. The findings highlight architecture-dependent sensitivity to initialisation and normalisation, providing practical guidance for deploying GNNs in AML pipelines. The study underscores the importance of training practices in GNNs and contributes to the development of more effective AML detection systems.

Key Points

  • ▸ GNNs are a principled approach to financial fraud detection, but their effectiveness depends on training practices such as weight initialisation and normalisation.
  • ▸ The study investigates the effectiveness of three GNN architectures (GCN, GAT, and GraphSAGE) with various initialisation and normalisation techniques.
  • ▸ The authors find that initialisation and normalisation are architecture-dependent and provide practical guidance for deploying GNNs in AML pipelines.

Merits

Strength in Empirical Analysis

The study provides a comprehensive empirical analysis of the impact of initialisation and normalisation on GNN performance, offering valuable insights for the development of more effective AML detection systems.

Practical Guidance for AML Pipelines

The study provides architecture-specific guidance for deploying GNNs in AML pipelines, which can help practitioners develop more effective detection systems.

Demerits

Limited Generalisability

The study focuses on a specific dataset (Elliptic Bitcoin) and may not be generalisable to other datasets or real-world scenarios.

Need for Further Research

While the study provides valuable insights, further research is needed to fully understand the impact of initialisation and normalisation on GNN performance and to explore other training practices.

Expert Commentary

The study provides a significant contribution to the development of more effective GNN-based anomaly detection systems, particularly in the context of AML. However, the findings may not be generalisable to other datasets or real-world scenarios, and further research is needed to fully understand the impact of initialisation and normalisation on GNN performance. The study's practical guidance for deploying GNNs in AML pipelines is particularly valuable, as it can help practitioners develop more effective detection systems. Nevertheless, the need for careful selection of initialisation and normalisation techniques highlights the complexity of GNN training and the importance of further research in this area.

Recommendations

  • ✓ Researchers should conduct further studies to explore the impact of initialisation and normalisation on GNN performance and to develop more effective training practices.
  • ✓ Practitioners should carefully select initialisation and normalisation techniques for GNNs based on the specific architecture and application, following the study's architecture-specific guidance.

Sources

Related Articles