Explainable fraud detection of financial statement data driven by two-layer knowledge graph

In modern economic activities, financial statement fraud will seriously mislead the economic decisions of investors and regulators, and will lead to huge investment losses even corporate bankruptcies. Although the powerful abilities have been gained by current machine learning methods in financial statement fraud detection problems, the explainability and the ability of extracting fraudulent patterns are still very scarce. In this study, an explainable Financial Statement data Fraud Detection method is proposed by introducing a Two-Layer Knowledge Graph (FSFD-TLKG) and a fraudulent pattern mining strategy on two-layer knowledge graph. Wherein, a two-layer knowledge graph comprises a semantic layer and a syntactic layer. Concretely, the subordination relationships among financial variables are represented in the semantic layer, and their articulation relationships are represented in the syntactic layer. Moreover, an explainable approach is designed to extract financial statement fraudulent patterns for credible fraud assertion. Experimental results show that, FSFD-TLKG can effectively extract explainable financial statement fraudulent patterns and obtain better detection accuracy than almost all traditional machine learning and deep learning methods except an unexplainable method: Extreme Gradient Boosting (XGBoost). Even for XGBoost, the explainable financial statement fraudulent patterns extracted by our method still can further improve its performance. Clearly, FSFD-TLKG gains the optimal practical performance which is much better than existing methods.