Multi-Dimensional Feature Analysis and Evaluation Methods for Anomalous Fund Flow Identification in Cross-Border Financial Transactions

Abstract

Cross-border financial transactions are inherently complicated by the multi-currency nature of the transaction, the different regulatory systems, and the various methods used to launder dirty money; anomaly detection is not easy. The paper presents a comprehensive, multi-sided, and characteristic-based analysis framework for anomaly detection in cross-border fund transactions. Using this framework by taking characteristics of transactions, network topology features, and temporal behavior patterns into account, in order to boost detection. A systematic evaluation was conducted on 2.8M transactional data; a combination of graph-structure-based features and time-series behavioral indicators outperformed a single-dimensional approach. After experimentation, this strategy increased the baseline's accuracy by 18.7 percent and recall by 23.4 percent.