Conditional Generative Models for Transparent Discrimination Detection and Mitigation in Algorithmic Decision-Making Systems

The increasing use of machine learning algorithms in automated decision-making systems has raised concerns about harmful and discriminatory decision patterns. While the literature on algorithmic fairness has focused on defining and eliminating discrimination, much fewer efforts have been made to detect it rigorously in black-box systems. State-of-the-art detection methods either translate notions of fairness into a statistical parity metric on the model’s output or consider additional knowledge about the structure of the world with a causal model. These approaches focus on equality of outcome and suffer from shortcomings, such as mutual incompatibility, lack of interpretability, and missing cases of intersectional discrimination. These shortcomings spur my research on complementary methods, which focus on equality of treatment by offering additional transparency into the sources of discrimination and being agnostic towards an a priori decision on the choice of protected features. In this project, I propose revealing a model’s decision-making process by leveraging recent advancements in generative models for tabular data. My contributions will increase transparency and reduce discrimination in many economic and societal applications, such as hiring, insurance, marketing, and mortgages. Finally, by understanding any black-box model’s internal logic, we can evaluate the discrimination cases without making additional statistical or causal assumptions that may bias the evaluation