Edge and Node Fusion for Transaction-level Prediction on Money Laundering Detection

An individual may have both legitimate and illicit transactions, each represented by distinct edges associated with a common node. Given that a node may encompass both legitimate and illicit transactions, we introduce an innovative edge and node fusion algorithm for transaction-level prediction on money laundering detection. Our algorithm boosts model performance by incorporating multi-task edge prediction and conditional edge prediction. Additionally, we propose a two-round training method to mitigate data imbalance by excluding high-confidence normal samples. The experimental results demonstrate the superiority of our algorithm in predicting illicit transaction edges compared to other baselines. Key findings of the study include significant improvements in recall and F1 score, highlighting the success of the two-round training method in enhancing model performance. Besides, the study provides insights into effective threshold selection, showcasing our algorithm's practical utility in combating money laundering activities.