DS Ph.D. Dissertation Defense | Kathleen Cachel | Wed. Jan. 8th @ 1:00PM, Salisbury Labs 105

DATA SCIENCE

PhD Dissertation Defense

Kathleen Cachel, Ph.D. Candidate, Data Science

Wednesday, January 8th, 1:00PM

Location: Salisbury Labs 105
 

Committee:

Elke Rundensteiner, PhD Advisor, Computer Science and Founding Head Data Science & AI

Lane Harrison, Computer Science and Data Science

Andrew Trapp, The Business School and Data Science 

Nicholas Mattei, Tulane University

Title: Fair Consensus Decision Making: Preference Aggregation Techniques for Candidate Group Fairness

Abstract:

This dissertation focuses on the preference aggregation problem of combining a diverse collection of voter preferences, called a preference profile, into a single consensus ranking representing voter preferences. Consensus rankings prioritize job applicants for employers, university students for funding, and patients for medical care. In each of these contexts, traditional strategies produce a consensus ranking without consideration for how this ranking may unfairly affect marginalized groups (i.e., race or gender). This dissertation explores the central question: How can we design preference aggregation strategies that are unbiased (fair) towards marginalized groups of ranked candidates while ensuring voter preferences are represented? We address this question through four tasks, each examining different types of preference data and their unique fairness concerns and bias mitigation needs.

In Task 1, we present the fair exposure kemeny aggregation problem, which combines fully ranked preferences into a consensus ranking that is fair toward candidate group. The goal is to ensure the consensus reflects voter preferences while ensuring candidate groups receive comparable amounts of exposure. We design both exact and approximate solutions to address this problem. Task 2 extends contemporary concepts of group fairness to the setting of non-conjoint ranked preference aggregation, where voter preference rankings may order different subsets of the full candidate set. To address this setting, we introduce WISE, a new plug-and-play approach for fair preference aggregation. Task 3 addresses the problem of fair partial ranked preference aggregation in contexts where voters provide partial rankings, meaning some candidates are intentionally left unranked. For this problem, we introduce PreFAIR a novel approach featuring a unique preference inference mechanism and a new Single Transferable voting inspired aggregation method. Task 4 explores the integration of group fairness into the previously unexplored area of fairly aggregating rated preference lists. We present a methodology called FATE that contains the first rating group fairness metric, along with algorithms designed for unique sources of bias in rating aggregation.