Tax evasion risk management using a Hybrid Unsupervised Outlier Detection method

作者：

Highlights：

• A novel hybrid unsupervised outlier detection method (HUNOD) for tax evasion risk management is proposed.

• HUNOD combines clustering and representational learning to detect and internally validate outliers.

• HUNOD can be enhanced by relevant domain knowledge reflecting underlying economic context.

• HUNOD relies on explainable-by-design surrogate models to increase interpretability of its results.

• Experimental evaluation shows that HUNOD detects between 90% and 98% of internally validated outliers.

摘要

•A novel hybrid unsupervised outlier detection method (HUNOD) for tax evasion risk management is proposed.•HUNOD combines clustering and representational learning to detect and internally validate outliers.•HUNOD can be enhanced by relevant domain knowledge reflecting underlying economic context.•HUNOD relies on explainable-by-design surrogate models to increase interpretability of its results.•Experimental evaluation shows that HUNOD detects between 90% and 98% of internally validated outliers.

论文关键词：Tax evasion,Outlier detection,Unsupervised learning,Clustering,Representational learning,Explainable surrogate models

论文评审过程：Received 2 March 2021, Revised 23 August 2021, Accepted 11 December 2021, Available online 4 January 2022, Version of Record 11 January 2022.

论文官网地址：https://doi.org/10.1016/j.eswa.2021.116409