Основні поняття
A randomized method for post-processing rankings that improves fairness without requiring access to protected attribute information.
Анотація
The paper proposes a randomized method for post-processing rankings to improve their fairness. The key insights are:
Fairness in ranking is an important problem, with applications in areas like HR automation and recommender systems. However, two challenges arise: (i) the protected attribute may not be available in many applications, and (ii) there are multiple measures of fairness in rankings, and optimizing for a single measure may lead to unfairness with respect to other measures.
The authors propose using Mallows noise, a distance-based probability model, as a randomization mechanism to improve fairness in a way that is oblivious to the specific protected attributes. This addresses the challenge of fairness without demographics.
The authors also address the challenge of robustness by showing that their method is robust with respect to P-Fairness, a common fairness measure for rankings, while maintaining competitive efficiency with respect to Normalized Discounted Cumulative Gain (NDCG).
Through extensive numerical experiments, the authors demonstrate the effectiveness of their randomized method. They show that it outperforms previously proposed methods like ApproxMultiValuedIPF and DetConstSort in terms of fairness and efficiency trade-offs.
The authors also conduct experiments on the German Credit dataset, showing that their method can improve fairness with respect to unknown protected attributes, while maintaining reasonable utility.
Overall, the paper presents a novel and practical approach to ensuring fairness in ranking algorithms, which is an important problem with significant real-world impact.