融合序列訊息與圖結構之反洗錢異常行為分析

Anti-Money Laundering (AML) is a critical research area in Financial Technology (FinTech) focused on detecting suspicious financial activity. However, the rise of new transaction types has led to increasingly subtle and complex money laundering schemes, rendering traditional rule-based methods inadequate for both detection and generalization. While machine learning and deep learning offer a promising alternative, there are still many challenges. To address these challenges, we propose an AML prediction framework based on sequence-graph fusion. Its core innovation is the joint modeling of an account’s individual temporal behavior and its structural features within the transaction network. Our approach begins by decomposing each account’s transaction history into incoming and outgoing sequences, which are encoded via a dual-branch Gated Recurrent Unit (GRU) to capture nuanced temporal patterns. We then utilize a bidirectional attention-based graph convolutional layer that employs a difference-aware message-passing mechanism to process relationships in both forward and backward directions, learning the behavioral contrasts between connected accounts. Through the attention mechanism, the model adaptively fuses each node’s intrinsic features with the aggregated features from both forward and backward neighbors. To counteract the extreme class imbalance inherent in AML data, our framework incorporates class re-weighting and balanced sampling strategies. We validated our proposed method on a public AML dataset. The experimental results demonstrate that our approach achieves stable F1-scores under severely imbalanced datasets, significantly outperforming traditional baseline methods.