Source-linked AI summary
Quantifying the Impact of User Attention on Fair Group Representation in Ranked Lists
Piotr Sapiezynski, Wesley Zeng, Ronald E. Robertson, Alan Mislove, Christo Wilson
TL;DR
Fairness assessments of ranked lists need to account for unequal user attention, which varies by service and context. The paper introduces the Viable-Λ Test to model that attention and assess group representation; its audits show that the same ranking can appear biased for or against a protected group depending on the attention distribution.
Problem
Fairness metrics for ranked lists often assume fixed attention discounting, although users' attention varies with service interfaces and search context.
Source: §1 · p. 1Method
The Viable-Λ Test evaluates group representation by combining item alignments with a parameterized model of user attention.
Source: §1 · p. 2Results
The same ranking can appear biased either for or against a protected group depending on the attention distribution used.
Source: §8 · p. 9Takeaways & Limitations
Assessing a service's fairness requires modeling attention for that particular service, rather than judging the ranking alone.
Source: §8 · p. 9The test's results rely on an accurate human-attention model, and the examples consider only truncated geometric attention distributions.
Source: §6 · p. 9
Abstract
from arXiv · showhide
In this work, we introduce a novel metric for auditing group fairness in ranked lists. Our approach offers two benefits compared to the state of the art. First, we offer a blueprint for modeling of user attention. Rather than assuming a logarithmic loss in importance as a function of the rank, we can account for varying user behaviors through parametrization. For example, we expect a user to see more items during a viewing of a social media feed than when they inspect the results list of a single web search query. Second, we allow non-binary protected attributes to enable investigating inherently continuous attributes (\eg political alignment on the liberal to conservative spectrum) as well as to facilitate measurements across aggregated sets of search results, rather than separately for each result list. By combining these two elements into our metric, we are able to better address the human factors inherent in this problem. We measure the whole sociotechnical system, consisting of a ranking algorithm and individuals using it, instead of exclusively focusing on the ranking algorithm. Finally, we use our metric to perform three simulated fairness audits. We show that determining fairness of a ranked output necessitates knowledge (or a model) of the end-users of the particular service. Depending on their attention distribution function, a fixed ranking of results can appear biased both in favor and against a protected group.
1 INTRODUCTION
The paper introduces the Viable-Λ Test to assess group fairness in ranked results while accounting for user attention. It argues that attention must be modeled for the service and that conclusions about bias can change with the attention distribution.
1 INTRODUCTION
- Fairness research has often assumed logarithmic attention discounting, although its slow decay can make the last eight results collectively receive more attention than the first result in a 100-result list.Prior work also used geometric attention without examining the consequences of varying its steepness.Source: §1 · p. 1
- The Viable-Λ Test asks whether any parameterized user-attention distribution makes a ranking group-fair and identifies the parameters of such a distribution.If the fitted attention model does not match observed user attention for the service, the system does not achieve representational parity.Source: §1 · p. 2
- The paper extends the metric to assignment uncertainty, aggregated results, multiple protected classes, and continuous protected variables.It also presents a blueprint for incorporating list usage patterns into fairness measurement.Source: §1 · p. 2
- The authors evaluate the metric on three real-world services: resume search, a dating service, and web search.The reported results show that attention-function choice and parameterization can yield different conclusions about whether and how rankings are biased.Source: §1 · p. 2
- The authors caution that fairness cannot be determined with certainty without knowing or modeling the audited service's true user-attention distribution.They frame the work as a metric for service operators or internal auditors, rather than as an audit study.Source: §1 · p. 2
2 BACKGROUND
Background work covers fairness concepts, ranking methods, and audits of real search engines. A recurring challenge is that fairness metrics must account for unequal and context-dependent user attention.
2 BACKGROUND
- Fairness research distinguishes individual treatment from group parity, while algorithmic fairness work also studies disparate treatment, disparate impact, and disparate mistreatment.Disparate mistreatment refers to false positives and negatives being unequally distributed across subpopulations.Source: §2 · p. 2
- Ranked lists raise an order-effects challenge because people tend to attend more to items near the top.Fairness methods developed for classification do not necessarily generalize to ranking.Source: §2 · p. 2
- Prior group-fair ranking methods set representation constraints, but some do not distinguish orderings; other work addresses attention through exposure measures or repeated reshuffling.The approaches include randomized ranking, user-specified constraints, and individual equity across multiple realizations.Source: §2 · p. 2
- Search-engine audits have examined gender stereotypes, employment-related race and gender fairness, and political partisanship.Earlier audits used metrics including average representation, top-K representation, logarithmic discounting, and linear rank normalization.Source: §2 · p. 2
- The user-attention distribution may be unknown and can vary with interface design and search context, so ad hoc attention metrics may misrepresent users' attention.The background describes this as a recurring challenge for evaluating fair ranked outputs.Source: §2 · pp. 2–3
3 METHODS
The methods combine item-level group alignment, modeled user attention, a population estimate, and a distance threshold to assess representation in ranked results. The Viable-Λ Test evaluates whether a plausible attention parameterization can bring group exposure within the chosen fairness threshold.
3 METHODS
- The Viable-Λ Test combines research from algorithmic auditing, information retrieval, and human-computer interaction to measure group fairness in search results.Its assessment incorporates human factors rather than treating the ranking algorithm in isolation.Source: §3 · p. 3
3.1 Overview
- The overview specifies an alignment vector, attentional weight vector, population estimator, distance metric, and maximum allowable distance for each ranked list.Alignments can be discrete or continuous; attention weights model relative user attention and are drawn from a parameterized family of truncated distributions.Source: §3.1 · p. 3
3.2 The Viable-Λ Test for Representational Parity
- Group exposure is computed from item alignments and attention weights, then compared with the population estimate using a distance metric and threshold.The threshold is context-dependent and chosen in conjunction with the distance metric.Source: §3.2 · p. 3, §3.1 · p. 3
- A ranking is unfair if no attention parameterization within the specified domain brings its exposure within the maximum allowable distance of the population estimate.Any acceptable parameterization must also match reasonable expectations and data about actual user behavior.Source: §3.2 · p. 3
4 DESIGN CHOICES
Viable-Λ models attention and group alignment explicitly, then tests whether a ranking is fair under plausible user-attention parameters. Its design also makes fairness judgments depend on population estimates, statistical thresholds, and the parameter space considered.
4.1 The Alignment Vector LR
- Representing each result’s alignment as a probability distribution accommodates ambiguous, multiple, or continuously varying group membership.The paper’s example uses three gender classes: Male, Female, and Unknown.Source: §4.1 · p. 3
4.2 Attentional Weights WR
- Attentional weights model relative attention by rank, which varies with interface and use context; click data may be unavailable or imperfect, and eye-tracking results may not transfer across sites.The paper assumes a parameterized family of truncated discrete distributions, including geometric, log-series, and discrete Pareto choices.Source: §4.2 · p. 4
- Inverse-log weighting decays slowly and has no parameters for adapting to different user behavior; for short lists, the paper instead expects users to read all results.For n = 1000, the last 10 results can have aggregate influence comparable to the first under inverse-log scaling.Source: §4.2 · p. 4
4.3 Estimating Population Demographics ˆp
- Estimating population proportions from the vendor’s results avoids requiring an external demographic reference, but makes the estimate dependent on vendor-data integrity and any sampling bias.If the vendor’s knowledge is biased, the metric measures how well its ranking represents that potentially biased knowledge.Source: §4.3 · p. 5
4.4 Distance Metric d and the δmax Threshold
- The distance d compares the attention-weighted group distribution with the estimated population distribution, while δmax sets the acceptable statistical distance.The threshold may account for uncertainty from result-list size and, for randomized rankings, the number of search queries.Source: §4.4 · p. 5
4.5 Restricting the Domain of the Parameter Space D
- Restricting the parameter space D to realistic attention behavior can increase test power by excluding implausible parameter values.Small-scale experiments can estimate attention parameters and thereby improve the power of the Viable-Λ test.Source: §4.5 · p. 6
4.6 Generating Fair Ranked Lists
- Fair ranked lists can differ with both class imbalance and attention steepness: a minority result may move higher, while steeper attention can favor a block of majority results after it.The examples compare class ratios of 1:10 and 5:10 under λ = 0.1 and 0.5.Source: §4.6 · p. 6
5 CASE STUDIES
The case studies apply the Viable-Λ Test to hiring, dating, and web-search rankings, using a geometric attention model. Their findings illustrate that fairness assessments depend on assumptions about how users distribute attention.
5.1 Gender in Hiring
- The hiring analysis used 412 result lists of at least 100 items and retained candidates’ gender-alignment probabilities rather than omitting ambiguous profiles or assigning them to a binary class.The estimate of population alignment for each query was the mean alignment in the vendor’s data.Source: §5.1 · pp. 6–7
- In hiring searches, about 92% of rankings passed the 1Z threshold, but 6% appeared biased against women and 3% unfair to men regardless of attention distribution.All rankings were deemed fair at the 2Z threshold.Source: §5.1 · p. 7
- Across sampled attention parameters, hiring rankings underrepresented women more often than men, although the dataset’s overall gender ratio was balanced.The analysis assumed users viewed 10% to 50% of results on average.Source: §5.1 · p. 7
5.2 Race in Dating
- For the dating case, the analysis simplified inferred race alignment to Black versus non-Black profiles and compared rankings with the population of 4,407 unique results.The dataset comprised 672 repeatedly collected lists, each with 100 results; the study used the vendor’s inferred labels.Source: §5.2 · p. 7
- Using unique profiles as the population estimate, dating rankings overrepresented Black profiles: they were 13% of unique profiles but 16% of displayed results on average.For individual lists, most were considered fair across attention parameters; which group appeared underrepresented varied with the assumed attention distribution.Source: §5.2 · p. 8
- Aggregating dating rankings could yield fair representation even when individual lists were unfair; steeper attention distributions required more runs for aggregates to be considered fair.The proportion of fair aggregates increased with the number of included runs, regardless of the assumed attention distribution.Source: §5.2 · p. 8
5.3 Political Bias in Web Search
- The political-bias case applied attention-weighted domain bias scores to search results, examining three queries selected from a dataset of 1,443 searches.The domain scores ranged from -1 for exclusively Democratic sharing to 1 for exclusively Republican sharing; the analysis reported whether weighted bias was positive or negative.Source: §5.3 · p. 8
- Political-bias judgments varied with attention: “financial regulation” appeared neutral, while “obamacare continue” and “medicare reform” could appear partisan in either direction.For “obamacare continue,” steeper attention could make the list appear liberal-leaning; for “medicare reform,” a conservative top result could make it appear strongly conservative.Source: §5.3 · pp. 8–9, Figure 7 · p. 9
6 LIMITATIONS
The examples use a truncated geometric attention model, and the population estimator assumes the displayed top k items proportionally represent all potential results. Both choices constrain how the audit can be applied.
6 LIMITATIONS
- The examples consider only a truncated geometric family for WR, although the framework permits arbitrary families.The test depends on an accurate human-attention model, motivating further research into perception and SEME quantification.Source: §6 · p. 9
- Single-parameter WR may miss attention patterns shaped by pagination and result-page features.Pagination can cause disproportionate attention drop-off between pages, while search engines may insert direct answers or group results by type.Source: §6 · p. 9
- The population estimator assumes the vendor’s top k items proportionally represent all N potential results, an assumption that may fail.If only the top 100 of 1,000 candidates are observed in a system that ranks women lower, the estimated female share may be too low.Source: §6 · p. 9
7 DISCUSSION
Changing rankings based on a measured attention distribution does not guarantee fair rankings, because users may change their behavior in response.
7 DISCUSSION
- Users’ attention may shift after results are reshuffled, so measured parity under one Λ does not guarantee fair rankings after changes.The operator could iteratively reshuffle for parity under measured Λ, re-estimate attention, and repeat.Source: §7 · p. 9
8 CONCLUSION
The conclusion presents a group-fairness metric tied to consumer attention and reports three real-world audits. It emphasizes that the same results can appear biased in opposite directions under different attention distributions.
8 CONCLUSION
- Across three real-world examples, the same results can appear biased for or against a protected group depending on the service’s attention distribution.The metric ties group-fairness measurement to consumers’ attention.Source: §8 · p. 9