Given the anonymity and complexity of illegal transactions, traditional deep-learning methods struggle to establish correlations between transaction addresses, cash flows, and physical users. Additionally, the limited number of labels for illegal transactions results in severe class imbalance and other challenges. To overcome these limitations, we propose a reinforcement learning-enhanced, multi-relational, attention graph-aware framework to detect anti-money laundering and illegal trading activities. On the one hand, a data-driven, graph-aware layer establishes long-term dependencies and correlations between transaction graph nodes. Similarity among graph nodes divides the topological graph into three subgraphs. Learning from these subgraphs and converging nodes enriches local, global, and contextual details. Simultaneously, using repeated nodes across the subgraphs enhances interactivity between them, reduces intra-class ambiguity, and accentuates inter-class differences. On the other hand, a reinforcement learning module embedded in the graph-aware layer compensates for the missing details in node features caused by masking operations. Furthermore, the reconstructed loss function addresses significant classification inaccuracies by reducing the weight assigned to easily classified samples. Balancing these issues and individually supervising each component enables the detection framework to achieve optimal performance. The evaluation results demonstrate that our proposed model exhibits optimal detection performance and robustness, such as F1 of 93.85% and 94.39%.

鉴于非法交易的匿名性和复杂性，传统的深度学习方法难以在交易地址、现金流和物理用户之间建立关联。此外，非法交易的标签数量有限，会导致严重的类别不平衡和其他挑战。为了克服这些限制，我们提出了一个强化学习增强的、多关系的、注意力图感知的框架来检测反洗钱和非法交易活动。一方面，数据驱动的图形感知层在事务图形节点之间建立了长期的依赖关系和相关性。图节点之间的相似性将拓扑图分为三个子图。从这些子图和收敛节点中学习可以丰富局部、全局和上下文细节。同时，在子图中使用重复的节点可以增强它们之间的交互性，减少类内歧义，并突出类间差异。另一方面，嵌入在图感知层中的强化学习模块可以补偿由掩码操作导致的节点特征中缺失的细节。此外，重建的损失函数通过减少分配给易于分类样本的权重来解决重大的分类不准确问题。平衡这些问题并单独监督每个组件使检测框架能够实现最佳性能。评估结果表明，我们提出的模型表现出最佳的检测性能和鲁棒性，例如 F1 分别为 93.85% 和 94.39%。