FA*IR: A Fair Top-k Ranking Algorithm

By Meike Zehlike, Francesco Bonchi, Carlos Castillo, Sara Hajian, Ricardo A. Baeza-Yates. 2017

547 citations recorded by the source corpus.

In this work, we define and solve the Fair Top-k Ranking problem, in which we want to determine a subset of k candidates from a large pool of n » k candidates, maximizing utility (i.e., select the "best" candidates) subject to group fairness criteria. Our ranked group fairness definition extends group fairness using the standard notion of protected groups and is based on ensuring that the proportion of protected candidates in every prefix of the top-k ranking remains statistically above or indistinguishable from a given minimum. Utility is operationalized in two ways: (i) every candidate included in the top-k should be more qualified than every candidate not included; and (ii) for every pair of candidates in the top-k, the more qualified candidate should be ranked above. An efficient algorithm is presented for producing the Fair Top-k Ranking, and tested experimentally on existing datasets as well as new datasets released with this paper, showing that our approach yields small distortions with respect to rankings that maximize utility without considering fairness criteria. To the best of our knowledge, this is the first algorithm grounded in statistical tests that can mitigate biases in the representation of an under-represented group along a ranked list.

Subfields: Artificial Intelligence, Economics and Econometrics, Safety Research

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Economics and Econometrics

In Safety Research

Citations in the corpus

FairSight: Visual Analytics for Fairness in Decision Making (2019, methodology): “We also include the in-processing model (Additive Counterfactual Fairness (ACF)) [26]) and the post-processing method (FA*IR [41]) as a way of mitigating bias ( R3 ).”
Fairness in Machine Learning: Lessons from Political Philosophy (2017, background): “…group impacts and treating like people alike do not apply; instead, the goal might be to ensure equal representation of groups in a ranking Zehlike et al. (2017), or to give due weight to different normative / ideological outlooks in a clas-siﬁer which automates the enforcement of norms…”
A Nutritional Label for Rankings (2018, methodology): “view of the Fairness widget in our example presents the output of three fairness measures: FA*IR [14], proportion [15], and our own pairwise measure.”
Investigating Societal Biases in a Poetry Composition System (2020, background): “More recently, researchers have proposed different metrics to measure the amount of group and individual fairness in ranking systems (Zehlike et al., 2017; Biega et al., 2018; Singh and Joachims, 2018; Yang et al., 2020).”
Bias in Bios: A Case Study of Semantic Representation Bias in a High-Stakes Setting (2019, background): “ 33]. Finally, although our paper does not directly consider ranking scenarios, fairness in ranking is particularly relevant to discussions about gender bias in online recruiting and automated hiring [7, 14, 22, 40, 42]. We quantify gender bias by computing the TPR gender gap— i.e., the difference i”
Using Image Fairness Representations in Diversity-Based Re-ranking for Recommendations (2018, background): “ learn hierarchical representations with multiple levels of abstraction [10]. Various e‡orts in incorporating fairness into ranking and recommendations have previously been pursued. In Zehlike et al. [21], the authors de•ned a group fairness criterion and an e†cient standalone algorithm for producin”
Algorithmic content moderation: Technical and political challenges in the automation of platform governance (2020, background): “…ranking, or blocking, may be more or less favourable to content associated with gender, race and other protected categories (Blodgett et al., 2016; Ekstrand et al, 2018; Zehlike et al., 2017), and thus entrench forms of representational harm against such groups (Barocas et al., 2017; Binns, 2018).”
Equity of Attention: Amortizing Individual Fairness in Rankings (2018, methodology): “This goal is achieved by techniques similar to those for ranking result diversification [6, 34, 36], or by granting equal ranking exposure to groups [30].”
Fairness-Aware Ranking in Search & Recommendation Systems with Application to LinkedIn Talent Search (2019, methodology): “Algorithms for ranking in the presence of fairness constraints, specified as the maximum (or minimum) number of elements of each class that can appear at any position in the ranking, have been proposed in [15, 42].”
Fairness in Recommendation Ranking through Pairwise Comparisons (2019, background): “In contract, [52, 44, 45, 11] explore what is a fair ranking, but focus on unpersonalized rankings where relevancy is known for all items”
Bias and Debias in Recommender System: A Survey and Future Directions (2020, methodology): “Towards group fairness, FA∗IR [117] is a post-processing method to achieve fair top-K ranking w.”
Fairness of Exposure in Rankings (2018, methodology): “ranking (top-10, top-20 and so on). The differences are then averaged for these prefixes using a discounted weighting (like in DCG). This measure is then used as a regularization term. Zehlike et al. [35] formulate the problem of finding a ‘Fair Top-k ranking’ that optimizes utility while satisfying”
Optimizing Long-term Social Welfare in Recommender Systems: A Constrained Matching Approach (2020, background): “Studies of fairness in ranking typically consider notions of group fairness , for example, by regularizing standard ranking metrics to encourage parity of the ranks across demographic groups (Yang & Stoyanovich, 2016; Zehlike et al., 2017).”
Gendered Language in Resumes and its Implications for Algorithmic Bias in Hiring (2022, background): “This calls for active consideration of fairness in downstream tasks such as employing “fairness through awareness” techniques (Dwork et al., 2012; Geyik et al., 2019; Zehlike et al., 2017) that explicitly take into account the protected class to achieve individual or group-level fairness.”
Mitigating bias in algorithmic hiring: evaluating claims and practices (2019, methodology): “In the computer science literature, there are ongoing efforts to define technical constraints on rankings in the spirit of equal representation and the 4/5 rule [18, 91, 94], and LinkedIn has adopted a similar approach to encourage demographic diversity in its search results [41].”