Model interpretability of financial fraud detection by group SHAP

作者：

Highlights：

• Our study compensates for the lack of interpretability of the financial AI black-box model.

• For the first time, marginal contribution measurement is employed to solve the financial fraud detection problem.

• Group SHAP managed to reflect the company’s abilities by evaluating group-wise features.

• Group SHAP greatly reduces computation time in Shapley value calculation.

• Our study accurately pointed out characteristics of different industries to meet multiple business needs.

摘要

•Our study compensates for the lack of interpretability of the financial AI black-box model.•For the first time, marginal contribution measurement is employed to solve the financial fraud detection problem.•Group SHAP managed to reflect the company’s abilities by evaluating group-wise features.•Group SHAP greatly reduces computation time in Shapley value calculation.•Our study accurately pointed out characteristics of different industries to meet multiple business needs.

论文关键词：Financial fraud detection,Model interpretability,Shapley value,Grouping,Kernel method

论文评审过程：Received 6 April 2022, Revised 31 July 2022, Accepted 1 August 2022, Available online 6 August 2022, Version of Record 13 August 2022.

论文官网地址：https://doi.org/10.1016/j.eswa.2022.118354