Equity of Attention

Abstract

Rankings of people and items are at the heart of selection-making, match-making, and recommender systems, ranging from employment sites to sharing economy platforms. As ranking positions influence the amount of attention the ranked subjects receive, biases in rankings can lead to unfair distribution of opportunities and resources such as jobs or income.This paper proposes new measures and mechanisms to quantify and mitigate unfairness from a bias inherent to all rankings, namely, the position bias which leads to disproportionately less attention being paid to low-ranked subjects. Our approach differs from recent fair ranking approaches in two important ways. First, existing works measure unfairness at the level of subject groups while our measures capture unfairness at the level of individual subjects, and as such subsume group unfairness. Second, as no single ranking can achieve individual attention fairness, we propose a novel mechanism that achieves amortized fairness, where attention accumulated across a series of rankings is proportional to accumulated relevance.We formulate the challenge of achieving amortized individual fairness subject to constraints on ranking quality as an online optimization problem and show that it can be solved as an integer linear program. Our experimental evaluation reveals that unfair attention distribution in rankings can be substantial, and demonstrates that our method can improve individual fairness while retaining high ranking quality.

The NodeHopper: Enabling Low Latency Ranking with Constraints via a Fast Dual Solver

“…The literature on learning for ranking problems with constraints can be organized into the following categories: Learning and ranking with specific constraints: There has also been work on ranking with specific classes of constraints like diversity [17] or fairness [4,7]. However, these approaches do not generalize to arbitrary constraints and rely on specific assumptions about the structure of the constraints.…”

“…While all the use cases described in [19] have this property, their approach can deal with different attention weights while ours depends on them being the same. The key contribution we make is that when this special structure is present, we can develop very efficient specialized solvers that enable significant speed-ups compared to passing (4) to an off-the-shelf LP solver.…”

“…It follows from duality theory for convex optimization [5] that (6) and (4) have the same solution and that a solution to (4) can be recovered from a solution to (6). We note a few advantages of the dual optimization problem: 1 The number of variables in (2) is n 2 (since the variable being solved for is an n × n doubly stochastic matrix) while in (6) it is just m. For most recommender systems we would have m << n and hence this is a huge gain in terms of computational complexity.…”

Fairness and discrimination in recommendation and retrieval

“…The issues of fairness, accountability, transparency, bias, discrimination, justice, and ethics that are seeing increased attention in many areas of computing also have signifcant relevance to the information retrieval community [3,8,9,12]. There is a substantial and rapidly-growing research literature studying fairness, bias, and discrimination in general machine learning contexts [5].…”

“…There is a substantial and rapidly-growing research literature studying fairness, bias, and discrimination in general machine learning contexts [5]. While some of this work, particularly work on fair ranking [3,15], translates easily into recommender and information retrieval systems, other issues such as the multisided nature of information discovery platforms [4] and the extreme sparsity of relevance judgments make it more difcult to apply fairness results from other felds to retrieval and recommendation settings.…”