What's in a Query: Polarity-Aware Distribution-Based Fair Ranking

TL;DR

This work introduces DistFaiR, a divergence-based, distribution-aware framework for amortized fair ranking that accounts for the full distributions of cumulative attention and relevance across a sequence of queries, and crucially incorporates query polarity to prevent fairwashing. By defining both individual and group unfairness via divergences between attention and relevance distributions, and proving that group unfairness is upper-bounded by individual unfairness for a broad class of divergences, DistFaiR provides a principled path to improve fairness without sacrificing performance. The authors operationalize this through an ILP-based re-ranking that minimizes the worst-case divergence while maintaining DCG-based quality, and they validate the approach across synthetic and real datasets, showing that polarity-aware methods yield stronger fairness improvements and reveal fairness-utility tradeoffs. The work highlights the practical importance of incorporating query properties, demonstrates the potential for improved fairness at both individual and group levels, and discusses risks of fairwashing, offering a concrete, testable framework for polarity-aware fair ranking in safety-critical settings.

Abstract

Machine learning-driven rankings, where individuals (or items) are ranked in response to a query, mediate search exposure or attention in a variety of safety-critical settings. Thus, it is important to ensure that such rankings are fair. Under the goal of equal opportunity, attention allocated to an individual on a ranking interface should be proportional to their relevance across search queries. In this work, we examine amortized fair ranking -- where relevance and attention are cumulated over a sequence of user queries to make fair ranking more feasible in practice. Unlike prior methods that operate on expected amortized attention for each individual, we define new divergence-based measures for attention distribution-based fairness in ranking (DistFaiR), characterizing unfairness as the divergence between the distribution of attention and relevance corresponding to an individual over time. This allows us to propose new definitions of unfairness, which are more reliable at test time. Second, we prove that group fairness is upper-bounded by individual fairness under this definition for a useful class of divergence measures, and experimentally show that maximizing individual fairness through an integer linear programming-based optimization is often beneficial to group fairness. Lastly, we find that prior research in amortized fair ranking ignores critical information about queries, potentially leading to a fairwashing risk in practice by making rankings appear more fair than they actually are.

Figures (8)

• Figure 1: Past work in amortized fair ranking minimizes the differences between an individual's expected cumulative attention and relevance over queries, where queries are considered exchangeable. However, such formulations lack critical information about the distributions of attention and the properties of queries (e.g., polarity). Our approach, DistFaiR, aims to overcome this. The example here considers two search queries with opposite polarities. Both individuals are equally relevant, and have equal expected attention, but have different attention distributions in (a) and (b).
• Figure 2: (a) and (b) show the difference (as relative change) between fairness metrics measured with and without query polarity. Query polarity impacts all amortized fairness metrics, as they differ from zero as seen in the plots. (rightmost) We plot the re-ranking performance of polarity agnostic and aware re-rankings under different permissible performance loss changes for the synth-cont dataset (DistFaiR($L_1$)), where we can see polarity agnostic re-ranking underperforms polarity aware re-ranking.
• Figure 3: Critical information about the distributions of relevance and attention (e.g., the variance) may be missing in such formulations.
• Figure 4: Past work in amortized fair ranking has ignored the impact of query polarity, and only considered expected cumulative attention. Here, if all individuals are equally relevant, and expected attention scores for ranks 1,2,3 are $\{0.5,0.5,0\}$ respectively, the sequence of queries appear fair because an individual's expected attention accumulated over the four queries is proportional to their relevance. However, we the female doctor is allocated attention only for the queries with negative polarity ("rude","bad dentist"). This leads to fairwashing.
• Figure 5: Individual Fairness Bounds Group Fairness under DistFaiR (here, $DistFaiR(W_1)$)

Theorems & Definitions (13)

• Theorem 3.1
• Remark 3.2
• Definition 4.1: DistFaiR-Divergence
• Lemma 4.2
• Definition 4.3: Amortized Individual Unfairness
• Definition 4.4: Amortized Group Unfairness
• Theorem 4.5
• Theorem 5.1
• Remark 5.2
• Theorem E.1