DualFraud: Dual-Target Fraud Detection and Explanation in Supply Chain Finance Across Heterogeneous Graphs

Bin Wu, Kuo-Ming Chao, Yinsheng Li

Abstract: In supply chain finance, detecting fraudulent enterprises and transactions is crucial to minimize financial loss. Enterprises and transactions have heterogeneous information and fraud labels, thus, leveraging such information well can simultaneously improve fraud detection performance in enterprise and transaction domains. This paper describes our newly proposed multitask learning framework, DualFraud, which detects fraudulent enterprises and transactions with explainability based on heterogeneous graphs in supply chain finance. The main contributions of this work are the proposed framework that can facilitate these two domains to share and enhance learning and modelling capabilities to improve fraud projection. The explainer component is attached to generate rich and meaningful explanations for risk controllers across enterprise and transaction graphs. Experiments on datasets prove the effectiveness of fraud detection and explainability in both enterprises and transactions. The proposed DualFraud outperforms the other methods in the selected criteria.