Scholarlink

<p>Detecting money laundering in cryptocurrency networks requires understanding complex patterns across sparse transaction graphs. Current graph neural networks face two key challenges. First, their layered design often causes over-smoothing: as information passes through multiple layers, node representations become too similar to distinguish. Second, these models fail to capture the irregular timing of illicit transactions, which typically occur in bursts rather than as smooth sequences. We introduce TGDAMFormer, a new approach that tackles both issues. Our method combines graph diffusion with arithmetic feature interactions to preserve the unique characteristics of distant nodes, thereby mitigating over-smoothing. To handle large-scale Bitcoin graphs efficiently, we pre-compute neighbourhood information before training, removing heavy computations from the learning process. For temporal patterns, we treat time as discrete periods rather than a continuous sequence. This allows the model to recognise laundering behaviours specific to certain time windows without relying on recurrent structures that struggle with sparse financial data. Experiments on the Elliptic++ dataset show that our model achieves state-of-the-art results. Using the Mean Average Distance metric, we demonstrate that node representations remain distinct even at depth (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\varvec{k}=\varvec{3}\)</EquationSource> </InlineEquation>), confirming that the over-smoothing problem is effectively alleviated. Importantly, we repurpose arithmetic interactions not merely as classifiers but as structure-aware filters that maintain feature diversity during graph propagation. Further tests on static benchmarks (TFinance and Yelp) show that these advantages extend beyond cryptocurrency analysis to general graph learning tasks.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Graph diffusion arithmetic feature interaction transformer with temporal information for bitcoin anti-money laundering

  • 期刊论文

  • 发表日期：

    2026年4月9日

查看全文
  • Meng Li,
  • Lu Jia,
  • Xinyi Ma,
  • Xinqiao Su

摘要

Detecting money laundering in cryptocurrency networks requires understanding complex patterns across sparse transaction graphs. Current graph neural networks face two key challenges. First, their layered design often causes over-smoothing: as information passes through multiple layers, node representations become too similar to distinguish. Second, these models fail to capture the irregular timing of illicit transactions, which typically occur in bursts rather than as smooth sequences. We introduce TGDAMFormer, a new approach that tackles both issues. Our method combines graph diffusion with arithmetic feature interactions to preserve the unique characteristics of distant nodes, thereby mitigating over-smoothing. To handle large-scale Bitcoin graphs efficiently, we pre-compute neighbourhood information before training, removing heavy computations from the learning process. For temporal patterns, we treat time as discrete periods rather than a continuous sequence. This allows the model to recognise laundering behaviours specific to certain time windows without relying on recurrent structures that struggle with sparse financial data. Experiments on the Elliptic++ dataset show that our model achieves state-of-the-art results. Using the Mean Average Distance metric, we demonstrate that node representations remain distinct even at depth ( \(\varvec{k}=\varvec{3}\) ), confirming that the over-smoothing problem is effectively alleviated. Importantly, we repurpose arithmetic interactions not merely as classifiers but as structure-aware filters that maintain feature diversity during graph propagation. Further tests on static benchmarks (TFinance and Yelp) show that these advantages extend beyond cryptocurrency analysis to general graph learning tasks.