Towards a Theory of Fairness in Machine Learning

Abstract:  Algorithm design has moved from being a tool used exclusively for designing systems to one used to present people with personalized content, advertisements, and other economic opportunities. Massive amounts of information is recorded about people’s online behavior including the websites they visit, the advertisements they click on, their search history, and their IP address. Algorithms then use this information for many purposes: to choose which prices to quote individuals for airline tickets, which advertisements to show them, and even which news stories to promote. These systems create new challenges for algorithm design. When a person’s behavior influences the prices they may face in the future, they may have a strong incentive to modify their behavior to improve their long-term utility; therefore, these algorithms’ performance should be resilient to strategic manipulation. Furthermore, when an algorithm makes choices that affect people’s everyday lives, the effects of these choices raise ethical concerns such as whether the algorithm’s behavior violates individuals’ privacy or whether the algorithm treats people fairly.

Machine learning algorithms in particular have received much attention for exhibiting bias, or unfairness, in a large number of contexts. In this talk, I will describe my recent work on developing a definition of fairness for machine learning. One definition of fairness, encoding the notion of ‘fair equality of opportunity’, informally, states that if one person has higher expected quality than another person, the higher quality person should be given at least as much opportunity as the lower quality person. I will present a result characterizing the performance degradation of algorithms, which satisfy this condition in the contextual bandits setting. To complement these theoretical results, I then present the results of several empirical evaluations of fair algorithms.

I will also briefly describe my work on designing algorithms whose performance guarantees are resilient to strategic manipulation of their inputs, and machine learning for optimal auction design.

Bio: Jamie Morgenstern is a Warren Center postdoctoral fellow in Computer Science and Economics at the University of Pennsylvania. She received her Ph.D. in Computer Science from Carnegie Mellon University in 2015, and her B.S. in Computer Science and B.A. in Mathematics from the University of Chicago in 2010. Her research focuses on machine learning for mechanism design, fairness in machine learning, and algorithmic game theory. She received a Microsoft Women’s Research Scholarship, an NSF Graduate Research Fellowship, and a Simons Award for Graduate Students in Theoretical Computer Science.