SHAP-Guided Risk Path Generation for IRS Auditors: Interpretable Anomaly Attribution in Payroll Tax Compliance Screening

Small and medium-sized enterprises are prone to errors or evasion in areas such as medical insurance reimbursement, pension contributions, and unemployment insurance, which not only harm workers' rights but also impact the sustainability of federal and state public funds. Therefore, an explainable intelligent tax compliance screening system is urgently needed. This paper uses multi-task learning as the backbone, while modeling the shared risk representations of sub-tasks such as wages, pensions, unemployment insurance, and medical insurance contributions. Subsequently, our model combines multi-task learning with time-series anomaly detection to capture overlooked scenarios and structural avoidance in long-term trends. On heterogeneous time-series graphs, SHapley Additive exPlanations (SHAP) contributions are propagated and decomposed according to 'feature-module-time-task,' and combined with gated dependencies and consistency constraints to generate auditable risk paths, achieving an upgrade from anomaly scoring to evidence-chain visualization. Experiments show that this method improves both identification performance and interpretable localization efficiency.