Imperceptible Fairness in Top-k Ranking Algorithms

Abstract

The top-k ranking of people and items in algorithmic systems helps people extract information and make decisions from vast amounts of data, raising widespread concerns about the fairness of these systems. Recent studies have considered various fairness issues in these systems, including fairness with different definitions and fairness for different user groups. In this paper, we show that users can notice violations in these rankings, leading users to distrust these rankings and thereby preventing the achievement of fairness goals. In particular, most known mechanisms try to minimize the utility loss incurred to achieve fairness, and these attempts provide loopholes for users to detect ranking violations (called selection violations). We introduce a model called imperceptible fairness, which deals with selection violations and proposes a feasible solution. Our experiments show that selection violations are a practical concern for real datasets and that our algorithm can prevent these selection violations with little overhead. Compared to baseline algorithms, we reduce the probability of intra-selection violations from 11.52% to 0.32% and inter-selection violation probability from 10% to 9.4%.

Date of Award	2026
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Raymond Chi Wing WONG (Supervisor)