Closing the gap

Abstract

According to the most recent strategic plan for the United States Internal Revenue Service (IRS), high-income individuals are a primary contributor to the "tax gap," the difference between the amount of tax that should be collected and the amount of tax that actually is collected [1]. This case study addresses the use of machine learning and statistical analysis for the purpose of helping the IRS target high-income individuals engaging in abusive tax shelters. Kernel-based analysis of known abuse allows targeting individual taxpayers, while associative analysis allows targeting groups of taxpayers who appear to be participating in a tax shelter being promoted by a common financial advisor. Unlike many KDD applications that focus on classification or density estimation, this analysis task requires estimating risk, a weighted combination of both the likelihood of abuse and the potential revenue losses.