Title:
Online Learning with a View toward Fairness
Online Learning with a View toward Fairness
Author(s)
Salem, Jad
Advisor(s)
Gupta, Swati
Editor(s)
Collections
Supplementary to
Permanent Link
Abstract
The field of fairness in algorithmic decision-making has flourished over the past decade. However, the overwhelming majority of this research concerns offline decision-making, where decisions are made en masse. In practice, however, decisions are often made in real time. In this dissertation, I use the domains of hiring and pricing to explore the themes of uncertainty and fairness in online-decision-making.
The first line of inquiry I will discuss revolves around uncertainty in online decision-making. Specifically, I focus on the task of online applicant-screening where uncertainty (e.g., noise in evaluations, or knowledge that certain pairwise rankings may be corrupted due to bias) is captured by a partial order on the set of applicants. Modeling the screening process as a $k$-secretary problem with partial ordinal information, I prove tight lower and upper bounds on the competitive ratio (i.e., the worst-case expected performance ratio).
The second line of inquiry revolves around achieving comparative fairness (i.e., ensuring that similar individuals receive similar decisions) in online learning. The challenge here is that stringent fairness constraints can prevent learning, thus resulting in poor performance. I thus introduce a novel relaxation of comparative fairness which allows decisions to monotonically vary over time. Under this constraint, I provide bandit convex optimization algorithms which attain optimal regret for $N=1,2$ and sublinear regret for $N>2$.
Finally, I turn to the law: what is allowed in addressing fairness and bias in algorithmic decision-making? First, I discuss the legality of fairness interventions in hiring. In particular, I show that precedent supports the use of partial orders to account for uncertainty in hiring. Turning to dynamic pricing, I discuss open questions regarding (a) how privacy laws affect contextual pricing, and (b) how price gouging laws can be translated into temporal fairness constraints.
Sponsor
Date Issued
2023-04-27
Extent
Resource Type
Text
Resource Subtype
Dissertation