An Active Learning Approach for Collecting Tax Revenue

• Activity Number: 548 - Using Artificial Intelligence and Advanced Statistical Methods to Improve Official Statistics
• Type: Topic Contributed
• Date/Time: Thursday, August 6, 2020 : 1:00 PM to 2:50 PM
• Sponsor: Government Statistics Section
• Author(s): Kristen Altenburger and Brandon Anderson and Ben Chugg and Jacob Goldin and Daniel En-Wenn Ho* and Ahmad Qadri and Evelyn Smith
• Companies: Stanford University and Stanford University and Stanford University and Stanford University and Stanford Law School and Internal Revenue Service and University of Michigan
• Keywords: active learning; internal revenue service; machine learning; artificial intelligence

Abstract

We propose to develop improved active learning methods to modernize how the Internal Revenue Service (IRS) collects tax revenue. From an “explore-exploit” framework, “research audits” (based on random selection) can be conceived of as exploring for potential underpayment, while “enforcement audits” (based on predicted risk) can be conceived of as exploiting areas of suspected underpayment. Our project proposes an active learning framework to consolidate these audit processes and to explore the gains from doing so. Active learning is distinct from conventional (offline) machine learning where one trains a model on a fixed dataset at one point in time and where the training dataset includes all labels (or outcomes). We will discuss preliminary results and highlight how this project will have to develop new methods to overcome common simplifying assumptions of active learners while enabling fairness and transparency.