PaperlayerView exact revision on arXiv

Source-linked AI summary

Fairness in Information Access Systems

Michael D. Ekstrand, Anubrata Das, Robin Burke, Fernando Diaz

arXiv:2105.05779v4cs.IR

TL;DR

Fairness in information access raises challenges beyond fair classification, including ranked outputs, personalization, user response, and multiple stakeholders. This monograph maps the problem space and surveys its literature, while recommending contextualized fairness goals and acknowledging its Western and especially U.S. focus.

  • Problem

    Information access systems raise distinctive challenges for identifying and operationalizing fairness, including multistakeholder applications, ranked outputs, personalization, and user response.

    Source: Introduction · p. 6

  • Method

    The monograph introduces information access and algorithmic fairness, develops a taxonomy of fair information access, surveys research, and offers guidance for future work.

    Source: Introduction · p. 7, §8.1 · p. 92, §8.2 · p. 94

  • Results

    The authors integrate prior work into a prospective map of fair information access, while noting that the map remains incomplete.

    Source: §8.3 · p. 95

  • Takeaways & Limitations

    The authors recommend defining a specific fairness-related harm and selecting a metric that plausibly captures it, with the measurement justified in context.

    Source: §8.2 · p. 94

    The monograph acknowledges a Western, especially U.S., focus rooted in the authors’ North American legal and moral context.

    Source: §1.7 · p. 14

Abstract

from arXiv · show

Recommendation, information retrieval, and other information access systems pose unique challenges for investigating and applying the fairness and non-discrimination concepts that have been developed for studying other machine learning systems. While fair information access shares many commonalities with fair classification, the multistakeholder nature of information access applications, the rank-based problem setting, the centrality of personalization in many cases, and the role of user response complicate the problem of identifying precisely what types and operationalizations of fairness may be relevant, let alone measuring or promoting them. In this monograph, we present a taxonomy of the various dimensions of fair information access and survey the literature to date on this new and rapidly-growing topic. We preface this with brief introductions to information access and algorithmic fairness, to facilitate use of this work by scholars with experience in one (or neither) of these fields who wish to learn about their intersection. We conclude with several open problems in fair information access, along with some suggestions for how to approach research in this space.

List of Key Terms

The monograph frames fairness in information access as a normative concern shaped by ranked, personalized, multistakeholder systems. It introduces key concepts and illustrates how fairness questions arise across applications and other responsibility concerns.

Introduction

  • Information access systems retrieve or recommend relevant items for user needs, but fairness also concerns how benefits and harms are distributed among affected stakeholders.The monograph focuses on fairness-related harms in the transaction that provides results in response to a request; other harms, such as disinformation amplification, are outside its main scope.Source: Introduction · pp. 6–7

1.1 Abstracting Information Access

  • Information access systems encompass retrieval, recommendation, filtering, and some natural-language-processing applications, despite differences among their research communities.Ranked outputs, personalized relevance, repeated decisions, and multistakeholder structure create shared challenges for fairness.Source: §1.1 · pp. 7–8

1.2 A Brief History of Fairness

  • Fairness traditions include rights, compensation, reward, and fitness, with fitness linking information access to allocating items by their utility to users.The monograph also discusses anti-discrimination law, disparate impact, and intersectionality as relevant parts of fairness history.Source: §1.2 · pp. 8–9

1.3 Fairness and Bias

  • The monograph distinguishes normative fairness judgments from bias as a systematic deviation in data, models, measurements, or outputs.Some biases create fairness problems, some do not, and some may be introduced to address another bias.Source: §1.3 · p. 9

1.4 Fairness and Other Responsibility Concerns

  • Fairness connects to responsibility concerns including ethics, accountability, transparency, safety, and privacy, while the monograph focuses on fairness-related connections.The cited discussion links accountability to mechanisms for holding organizations responsible for societal fairness goals.Source: §1.4 · p. 10

1.5 Running Examples

  • Running examples ask whether systems fairly allocate exposure, service quality, opportunities, and representation across affected groups and users.Examples span music discovery, news, jobs, and philanthropic recommendations, including concerns about balanced news exposure and feedback loops that let some causes dominate.Source: §1.5 · pp. 11–12

1.6 How to Use Tis Monograph

The monograph is organized to orient readers from information access and algorithmic fairness toward a taxonomy, literature surveys, and future research. It emphasizes contextual problem framing while acknowledging its Western and especially U.S. perspective.

1.6 How to Use Tis Monograph

  • The monograph targets information-access researchers and algorithmic-fairness researchers, organizing background chapters before its taxonomy, stakeholder-focused surveys, temporal analysis, and future-work guidance.Readers are directed to chapters 2–4 and 8, then to later chapters relevant to their work.Source: §1.6 · p. 13

1.7 Our Perspective

  • The authors treat fairness as one family of system objectives, not a complete account of social or ethical goals, and urge application-specific problem framing.They argue that metrics and mitigations should be developed and assessed for a specific, well-defined problem, with relevant experts helping identify harms and objectives.Source: §1.7 · p. 14

1.8 Some Cautions

  • The authors acknowledge that their treatment has a Western and especially U.S. focus, and call for careful, compassionate work that avoids overbroad claims.They note that all four authors work in North America and approach the topic primarily in that legal and moral context.Source: §1.7 · p. 14, §1.8 · p. 14

Beware Abstraction Traps

Fairness research must attend to the social context of specific harms rather than assume abstractly similar problems share causes or solutions. The section also outlines common limits in fairness research and the information-access systems the monograph addresses.

Beware Abstraction Traps

  • Abstract problem forms can obscure social impacts, and similar discrimination can arise through distinct causes and mechanisms, so harms and interventions require context-specific validation.The authors note that general detection and mitigation methods may be useful, but their effectiveness needs revalidation for the particular harms they address.Source: Beware Abstraction Traps · p. 15

Beware Limits

  • Common research limits include single-dimensional or binary attributes, treating social categories as fixed, and incomplete, erroneous, or biased data.Single-binary-attribute studies can be useful starting points, but authors should contextualize them and account for known limitations.Source: Beware Limits · p. 16

Beware Convenience

  • Researchers should expand available resources rather than take the easiest path, and cite underlying studies for their findings rather than relying on this monograph alone.The authors welcome citations for their organization and synthesis of fair information access research.Source: Beware Convenience · p. 16

Information Access Fundamentals

  • Information access retrieves items from a repository to help satisfy a user’s information need, using explicit or implicit expressions of that need and varied forms of presentation and feedback.Presentations may be ranked lists or other structured outputs, and satisfaction may be measured explicitly or through behavior such as clicks, streams, or purchases.Source: Information Access Fundamentals · p. 17

2.1 System Overview

  • The typical pipeline connects item and user understanding with retrieval and rendering, then uses behavior understanding to inform retrieval and evaluate whether needs are met.The stages include understanding items and users, retrieving matching items, presenting them, and interpreting user behavior.Source: §2.1 · pp. 17–18, Figure 2.1 · p. 18

2.2 Repository of Items

  • Item repositories reflect provider creation and collection policies, and represent items through content, metadata, and usage data that can be biased by system decisions and updated over time.Usage data may include users who consumed an item or queries that led to clicks.Source: §2.2 · p. 19

2.3 Users and Information Needs

Information access systems represent needs in explicit and implicit ways, present retrieved items through different formats and interactions, and typically make ranked decisions. The section describes how item utility and system performance can be estimated and evaluated.

2.3 Users and Information Needs

  • Users’ needs may be explicit, expressed through features, examples, or language, or implicit, reflected in global user properties and local session context.The distinction between global and local implicit information can be fluid, especially across sessions with a shared goal.Source: §2.3 · p. 20

2.4 Presentation

  • Presentations range from a single item and ranked list to two- or three-dimensional layouts, sometimes with summaries that let users inspect items without consuming them fully.A system may also identify a specific content segment to help users make a decision.Source: §2.4 · p. 21

Interaction and Sessions

  • Information needs often span multiple interactions, with system decisions interleaved with user feedback that can guide adaptation across turns or sessions.Examples include search sessions, navigation through recommendation interfaces, and repeated single-item decisions with user skips.Source: Interaction and Sessions · p. 21

Rankings

  • Rankings are the fundamental algorithmic decisions in information access, even when systems rearrange ranked results into grids or combine multiple rankings.The most common ranking sorts by decreasing relevance under the probability ranking principle.Source: Rankings · pp. 21–22

2.5 Evaluation

  • Evaluation may be situated with real users or simulated using data and behavior models; simulation is repeatable and efficient but assumes it predicts performance in deployment.Situated evaluation is more costly in user risk, time, and engineering effort.Source: §2.5 · p. 22

Estimating Item Utility

  • Item utility can be estimated from explicit labels or implicit behavior, but label meaning and behavioral signals vary by domain and may not reflect longer-term value.When longer-term utility is measurable, methods such as causal inference, reinforcement learning, and multicriteria optimization can attribute it to earlier decisions.Source: Estimating Item Utility · pp. 22–23

Evaluating System Decisions

  • System-decision evaluation measures utility across ranked outputs using browsing models and rank discounts, with metrics encoding assumptions about which items users inspect.Offline evaluation simulates user behavior from utility data and system decisions; online evaluation can aggregate behavioral signals in A/B tests.Source: Simulated Evaluation · pp. 23–24, Situated Evaluation · p. 23
  • Rank-biased precision models exponentially decreasing inspection probability by rank, making its expected utility proportional to the last item inspected.Its browsing model includes the hyperparameter γ.Source: Simulated Evaluation · p. 24

Evaluation Encodes Human Values

Information access systems rank items using utility estimates, often shaped by learned or personalized models, while fairness requires attention to the values embedded in evaluation and to harms beyond aggregate effectiveness. The chapter surveys these foundations and stresses that fairness goals are contextual rather than universal.

Evaluation Encodes Human Values

  • Evaluation metrics and utility signals encode contestable personal and societal values, so designers should interrogate them before treating them as objective.This includes assessor guidelines, heuristics based on log data, and choices about longer-term utility.Source: Evaluation Encodes Human Values · p. 24

2.6 Algorithmic Foundations

  • Information access algorithms estimate item utility for an information need and use those estimates to produce deterministic or stochastic rankings.The chapter introduces these foundations selectively to explain how information access fairness differs from fairness in other machine-learning applications.Source: §2.6 · p. 25

Vector Space Model

  • Vector-space models represent documents and needs as term-based vectors, compare them for relevance, and can combine item content with user history.Collaborative filtering instead relies on historical user-item interactions; both document-term and ratings matrices are sparse and incomplete.Source: Vector Space Model · pp. 25–26

Embedding and Optimizing Utility

  • Latent embeddings provide compact representations that can match documents to needs through synonyms, while learned utility models estimate relevance from observations such as clicks or ratings.Modern models may use ratings, click-through probabilities, or relevance probabilities; optimization can learn utility directly.Source: Embedding and Optimizing Utility · pp. 26–27

User Modeling

  • Personalized systems represent a user’s preferences from interaction history and use that representation in scoring and ranking items.User models may distinguish persistent preferences from short-term interests.Source: User Modeling · p. 27

Learning to Rank

  • Learning to rank optimizes item ordering rather than the accuracy of pointwise utility estimates; pairwise methods can train ranking models across recommendation and retrieval.Bayesian Personalized Ranking optimizes the probability that a randomly selected relevant item scores above an irrelevant one.Source: Learning to Rank · pp. 27–28

Re-ranking

  • Re-ranking adjusts an initial ranking to meet additional goals, including diversity, or to apply a more computationally costly model to a smaller candidate pool.Maximum marginal relevance balances relevance against similarity to items already ranked.Source: Re-ranking · p. 28

Fairness Fundamentals

  • Algorithmic fairness examines distributions of effects and representation, but fairness concepts are contextual and contestable rather than one universal objective.The chapter surveys core concepts and sources of unfairness; identifying and addressing particular harms is application-specific and calls for stakeholder and subject-expert input.Source: Fairness Fundamentals · pp. 29–30

3.1 Sources of Unfairness

Unfairness can enter throughout the decision-support pipeline, from social conditions and data collection through modeling, evaluation, and human response. Because it can be reintroduced downstream, identifying a fairness problem requires specifying the harm, affected people, and normative aims.

3.1 Sources of Unfairness

  • Unfairness may arise at any pipeline stage and be propagated, mitigated, exacerbated, or reintroduced downstream, so different stages require different interventions and measurements.The loop runs from observations through model results and evaluation to human responses that affect the environment observed in later rounds.Source: §3.1 · pp. 31, 32, 34, Figure 3.1 · p. 32
  • Historical discrimination can make observed features differ across groups for reasons unrelated to relevant underlying ability, as redlining shaped access to wealth and neighborhood resources.The discussion distinguishes observed features from the ideal construct features relevant to a decision, such as ability to repay a loan.Source: §3.1 · pp. 31–33
  • Data collection can introduce unfairness through sampling, response, measurement, and variable-selection choices, as well as through the perspectives and categories used to define and record constructs.Loan-repayment-only data, for example, can exclude evidence about borrowers with little borrowing history; group size can also make naïve models less accurate for minority groups.Source: §3.1 · p. 33
  • Models can reproduce unfairness through direct sensitive-attribute use, proxy variables, or objectives that encode particular views of a good model.Source: §3.1 · p. 34
  • Evaluation choices determine whose utility counts; aggregate metrics prioritize majority-group performance, and historical lending records can omit outcomes for denied applicants.Disaggregating and reweighting metrics can favor systems with more equal performance across groups; missing repayment outcomes can leave false negatives unobserved.Source: §3.1 · p. 34

3.2 Problems and Concepts

  • Coherent fairness work must identify who is harmed, how harm manifests, and how it is measured, while making its normative principles, assumptions, and goals explicit.The chapter treats fairness as specific harms to assess and address in their application and domain context.Source: §3.2 · p. 35

Distribution and Representation

Fairness concepts distinguish harms and operational goals, but their application depends on assumptions about similarity, group parity, and the validity of observed outcomes. The chapter emphasizes that these choices carry limits and trade-offs.

Distribution and Representation

  • Distributional harms concern denial of resources or benefits, while representational harms concern inaccurate portrayals that can also affect decisions and resource allocation.The surveyed literature has focused mostly on distributional harms, often treating representational fairness as a means to reduce them.Source: Distribution and Representation · pp. 35–36

Individual Fairness

  • Individual fairness requires similar people to receive similar decisions, but depends on a task-specific fair distance metric and can reflect biased observed features rather than underlying similarity.Under WYSIWYG, observed features are assumed unbiased; when systemic discrimination affects observations, the guarantee applies only in that potentially biased observation space.Source: Individual Fairness · pp. 36–37

Group Fairness

  • Group fairness includes statistical parity, often assessed by comparing decision rates; the four-fifths rule flags a protected-group pass rate below 80% of the dominant-group rate.Parity measures can help detect potential discrimination, but are often insufficient evidence of discrimination or legal liability.Source: Disparate Impact and WAE · pp. 38–39
  • Error-based group metrics rely on WYSIWYG labels, yet equalizing multiple error rates can be impossible with unequal base rates and imperfect classifiers.The Chouldechova-Kleinberg result rules out simultaneously equalizing more than two error-parity metrics under those conditions; WAE assumes equal base rates but can entail biased labels.Source: Error-Based Constructs and WYSIWYG · pp. 39–40
  • Fairness measures also face partial-outcome data: lending repayment is observed only for issued loans, and information-access feedback is observed only for results shown to users.Meritocratic fairness prevents preferring a less-qualified candidate but does not bound how large the probability gap between candidates may be.Source: Error-Based Constructs and WYSIWYG · p. 40

Awareness, Treatment, and Impact

Fairness concepts and interventions depend on assumptions about protected attributes, social goals, and where inequity enters the system. The chapter surveys mitigation approaches from data preprocessing through postprocessing, while emphasizing that interventions require system-level evaluation.

Awareness, Treatment, and Impact

  • Ignoring protected-group status does not ensure fairness because other features can act as proxies, while different fairness constructs rely on distinct assumptions and definitions.Individual fairness focuses on task-relevant similarity; statistical parity compares positive-decision rates, and error parity compares group decision accuracy.Source: Awareness, Treatment, and Impact · pp. 40–41

Motivations and Teories

  • Fairness objectives reflect different assumptions and philosophies: anti-classification removes protected traits from decisions, while anti-subordination seeks to reverse the effects of past discrimination.The two approaches may share technical constructs but can lead to different end goals.Source: Motivations and Teories · p. 41

3.3 Mitigation Methods

  • Mitigation can target different pipeline stages, but fairness does not necessarily compose, so interventions require assessment of the full system in its actual use.An intervention need not occur at the source of the problem; downstream stages may reintroduce unfairness or create new forms.Source: §3.3 · p. 41

Preprocessing

  • Preprocessing methods include suppressing attributes, changing labels, reweighting examples, and stratified sampling; feature repair removes predictive correlations with sensitive attributes regardless of causal connections.Feldman et al. (2015) describe altering observed features to reduce their utility for predicting sensitive attributes.Source: Preprocessing · p. 42

Representation Learning

  • Fair representation learning seeks latent representations that reduce sensitive information while retaining other information, using approaches such as prototypes, matrix transformations, and adversarial objectives.It can be formulated as jointly minimizing information loss and removing sensitive-attribute information.Source: Representation Learning · pp. 42–43

Fairness-Aware Decision Models

  • Fairness-aware decision models incorporate fairness objectives into training, commonly through loss penalties or constraints, but these approaches can be unstable under small dataset changes.Adversarial learning is another approach; adaptive data collection can also iteratively combine modeling with gathering or preprocessing data.Source: Fairness-Aware Decision Models · p. 43

Postprocessing

  • Postprocessing leaves data and models unchanged while adjusting outputs, including group-specific thresholding; reranking is particularly important for fairness-aware information access.Thresholding can use different decision boundaries for groups to satisfy a fairness definition.Source: Postprocessing · pp. 43–44

3.4 Wrapping Up

Fairness in information access requires context-specific objectives because classifier fairness assumptions often fail in ranked, personalized, interactive settings. The chapter maps relevant harms and stakeholders, emphasizing that measures of user performance alone may miss inequities.

3.4 Wrapping Up

  • Fairness objectives and metrics are not universally compatible or transferable, so they must be selected for the application’s ethical, legal, and sociotechnical context.The chapter’s brief introduction points readers to further work on algorithmic fairness.Source: §3.4 · p. 44

Te Problem Space

  • Information access fairness spans diverse potential harms, some not yet documented in the wild, and the chapter’s taxonomy uses overlapping, non-hierarchical facets to organize them.The authors aim to guide proactive consideration of potential harms and position existing and future work.Source: Te Problem Space · pp. 45–46

4.1 What Breaks in Information Access?

  • Classifier fairness does not transfer directly because information access allocates rivalrous ranked positions, makes repeated and personalized decisions, and evaluates subjective outcomes across multiple stakeholders.Ranked decisions are interdependent, and repeated decisions can also be assessed in expectation or amortized across requests.Source: §4.1 · pp. 46, 48, 49, 50

4.2 Kinds of Harms in Information Access

  • Optimizing mean performance can improve aggregate metrics while under-serving under-represented groups, whose attrition may further affect consumers and providers in two-sided recommendation.In simulation experiments, Hashimoto et al. (2018) demonstrate these dynamics under traditional empirical risk minimization.Source: §4.2 · pp. 50–51

4.3 Fair for Who?

  • Information access systems affect multiple stakeholder groups, and different applications can require different distributions of utility and personalization across those groups.Consumers and providers are salient stakeholders, but fairness requirements can also be imposed by organizations, regulators, or interest groups.Source: §4.3 · pp. 51–52

Consumers

  • Consumer fairness includes service quality, content received, and the costs of using a system; these concerns need not align, so comparable performance metrics may hide unequal access to desirable inventory.Clicks can suggest satisfaction with shown results without revealing whether users would prefer results shown to another group; tracking and profiling may also feel more threatening to marginalized users.Source: Consumers · pp. 52–53

Providers

Fair information access concerns extend beyond consumers to providers, information subjects, and indirectly affected parties. The chapter distinguishes these harms and notes that fairness can involve multiple groups and forms of representation at once.

Providers

  • Provider fairness concerns the exposure of providers’ items, often across multiple visits because rankings are limited and personalized; evaluating it is complicated when items are poor fits for users.The chapter notes that little data is available on providers’ own fairness perceptions, and most research focuses on consumers.Source: Providers · pp. 53–54

Subjects

  • Subject fairness concerns representation of people depicted or discussed in retrieved items, including risks of stereotype reinforcement and gaps in medically relevant information.The authors note little significant research on the potential medical-information problem and call for proactive study.Source: Subjects · pp. 54–55

Side Stakeholders

  • People indirectly affected by recommendations or routing, such as delivery drivers and residents on diverted streets, may have fairness concerns despite not participating in the interaction.The chapter identifies comparatively little published work on these side stakeholders.Source: Side Stakeholders · p. 55

Joint Fairness

  • A system may need to address simultaneous fairness concerns across stakeholder groups and multiple protected groups, while intersectional challenges remain insufficiently addressed.The chapter illustrates this with rental listings affecting both renters and landlords, and image search representing women and ethnic minorities.Source: Joint Fairness · pp. 55–56

Cross-Group Harms

  • Unfairness affecting one group can also harm others, so evaluations and interventions must identify whose concerns they measure because groups may require different metrics and techniques.Provider- or subject-unfair systems may skew consumers’ perceptions, while consumer-unfair systems may discourage a provider’s primary audience.Source: Cross-Group Harms · p. 56

4.4 Fair How?

  • Fairness can concern distribution of exposure or information, as well as representational harms that misrepresent users, items, or the information space; categories may overlap and some harms are legally defined.Internal representation problems may lead to other harms or be compensated for by other system behavior.Source: §4.4 · pp. 56–57

Direct Misrepresentation

  • Direct misrepresentation can give consumers inaccurate information and make items harder to find, while systematic errors may misrepresent providers, groups, or broader communities.The chapter’s examples include inaccurate book labels, incomplete gender data for authors, and a search result that misrepresented Cherokee identity.Source: Direct Misrepresentation · pp. 57–58

Unfair Result Set Composition

Unfairness in information access can arise in result composition, the distribution of benefits, evaluation vantage points and time scales, and throughout the system pipeline. Its effects depend on which stakeholders and resources are involved, and can accumulate through user behavior and system learning.

Unfair Result Set Composition

  • Result composition can reinforce stereotypes and shape perceptions of who belongs in a role; personalized profiles or item representations can also produce stereotyped recommendations.The “CEO” image-search example illustrates effects on both users’ perceptions and the images available to them; movie recommendations can similarly reflect gender stereotypes.Source: Unfair Result Set Composition · p. 58

Unfair Distribution of Benefits

  • Distributional fairness concerns include unequal quality or exposure for similar users or content, and harms vary with whether the resource is subtractible.Ranking positions are scarce for providers, while consumer-side resources may be non-subtractible or scarce only when successfully consumed.Source: Unfair Distribution of Benefits · pp. 59–60

4.5 Fair from What Vantage Point?

  • Fairness can be assessed through prediction accuracy, objective result quality, or item appearance frequency; fairness concerns may also exceed quantitative exposure or utility.The relevant vantage point depends on the application and stakeholder, including whether consumers receive lower-quality items or providers receive less exposure.Source: §4.5 · p. 61

4.6 Fair on What Time Scale?

  • Point-in-time offline evaluation may miss fairness outcomes delivered over time, as popularity feedback loops can amplify unequal item exposure.Systems can assess a metric over a time window and adapt in a later interval, but repeated interaction and model updates mean point-in-time analysis is insufficient.Source: §4.6 · pp. 61–62

4.7 Fairness and the System Pipeline

  • Unfairness can enter through data or models at any pipeline stage, including item understanding, retrieval, rendering, behavior learning, and evaluation.Biased feedback can amplify small differences into large exposure gaps, while biased relevance judgments or response data can lead to unfair design or model choices.Source: §4.7 · pp. 62–64

4.8 Fairness and Other Concerns

Fairness overlaps with other information-access concerns but has distinct normative aims, and its relationship to accuracy or utility remains unsettled. The monograph connects information access to broader fairness research and organizes its survey around harms, while describing consumer fairness at both individual and group levels.

4.8 Fairness and Other Concerns

  • Fairness extends concerns about incidental item properties to biases linked to the broader social location of items, providers, users, and others.Diversity methods may be reused for fairness, but the two concerns have different normative foundations and require metrics reflecting fairness.Source: §4.8 · p. 64
  • The relationship between fairness and accuracy or utility may vary by setting, and the available research does not support general conclusions.The relationship between fairness and utility may differ from its relationship with offline accuracy.Source: §4.8 · p. 65

4.9 Contributing Back to ML Fairness

  • Information access offers a setting for examining classical fairness assumptions, with benchmark datasets and a different impact profile from criminal justice or lending.The authors note that transfer to other applications requires care, while information access may serve as a testing ground for techniques with broader applicability.Source: §4.9 · pp. 65–66

4.10 Navigating the Problem Space

  • The authors present their taxonomy as a useful, necessarily incomplete and imprecise way to organize harms, prioritizing clear harm descriptions over rigid category placement.They intend the framing to guide people building or affected by information access systems and to organize the monograph’s subsequent discussion.Source: §4.10 · p. 66

Consumer Fairness

  • Consumer fairness concerns whether a system’s effects on consumers and subgroups result in unjust harms.For example, job seekers in different groups may receive results of different quality.Source: consumer fairness · p. 67

5.1 Individual Fairness

  • Individual fairness can be examined through the distribution of utility across requests, but equalizing accuracy without context can harm users with stronger profiles.Comparisons can account for profile size, though the choice of similarity features itself raises normative questions.Source: §5.1 · p. 68

5.2 Group Fairness through Disaggregated Utility

  • Group fairness can be assessed by disaggregating utility across consumer groups; reported studies found group differences that changed after query matching or training-data resampling.Mehrotra, Anderson, et al. found age and gender satisfaction differences reduced almost to zero after context matching; Ekstrand, Tian, Azpiazu, et al. found significant nDCG differences that resampling to equal gender representation substantially reduced.Source: §5.2 · pp. 68–69

5.3 Disparate Effectiveness

Consumer-side differences in information-system effectiveness may arise from factors across the access pipeline, so identifying their sources matters before choosing a response. The discussion extends beyond accuracy to other consumer concerns, complex group settings, and provider fairness.

5.3 Disparate Effectiveness

  • Disparate effectiveness is a difference in how well an information system satisfies different groups’ needs; its causes may lie anywhere in the pipeline, including measurement, training data, relevance models, or mediating factors.Per-user or per-query utility can reveal such disparities, but the observed difference alone does not identify its cause.Source: §5.3 · p. 70
  • Mehrotra, Anderson, et al. (2017) found that context matching eliminated most disparity by controlling mediators, while Ekstrand, Tian, Azpiazu, et al. (2018) found training-data availability mattered but profile size did not.Matching can discard unmatched queries and obscure inequity if those queries are where the disparity arises; downsampling is useful diagnostically but is not necessarily advisable in applications.Source: §5.3 · pp. 70–71

5.4 Providing Fair Utility

  • Because user-side quality is not usually rivalrous, reducing better-served users’ results is not an obvious remedy; improving service for difficult underserved queries may address disparities without harming others.The authors recommend caution with regularizers aimed simply at reducing consumer-side effectiveness disparities.Source: §5.4 · pp. 71–72

5.5 Fairness Beyond Accuracy

  • Consumer fairness also concerns discriminatory item or ad exposure, recommendations independent of protected class, and representations designed to reduce sensitive-attribute predictability.The cited work includes ad-distribution disparities arising from relevance optimization and market dynamics, and approaches to independent recommendations and fair embeddings.Source: §5.5 · p. 72

5.6 More Complex Scenarios

  • Consumer-fairness research often treats protected attributes separately and as binary, leaving compound intersectional disadvantage and continuous sensitive attributes largely unaddressed in recommendation fairness.The monograph notes that no existing work addresses the compound disadvantage at identity intersections it describes.Source: §5.6 · p. 73

Provider Fairness

  • Provider fairness concerns benefits content suppliers gain when systems connect users with their items; one item can have multiple providers, and diversity addresses consumer intent rather than justice.Providers may include creators, publishers, labels, or other contributors; provider fairness focuses on fair opportunity for content discovery.Source: Provider Fairness · pp. 74–75

6.1 Provider Representation

  • Provider representation measures how result lists represent provider groups; skew may create representational harm or signal denied exposure, and these constructs are usually measured separately from relevance.The concern is the system’s portrayal of the provider population, not an individual provider’s internal or external representation.Source: §6.1 · p. 76

Measuring Representation

Representation-based fairness compares provider-group distributions in result lists with a target distribution, but the target and treatment of rank positions matter. The cited studies examine gender representation, fair re-ranking, and newer exposure-based approaches.

Measuring Representation

  • Representation measurement combines a target distribution over provider groups, a distance function, and a way to compute and compare the list’s group distribution; the appropriate target depends on domain and fairness goals.Possible targets include a uniform distribution, the provider population, relevant providers, or a societal estimate.Source: Measuring Representation · pp. 76–78
  • Overall list composition ignores positional visibility; prefix tests and rank-discounted distributions incorporate the fact that higher-ranked items may matter more.Attention-Weighted Rank Fairness supports more than two groups and can accommodate real-valued or mixed group-membership weights.Source: Measuring Representation · pp. 77–78

Studies of Gender Representation

  • Book recommendations reflected users’ author-gender biases; Spotify streams favored male artists, while female programmed-stream share correlated with greater non-programmed listening.Book users tended to read more men, with diffuse gender tendencies; Spotify analysis found no age or gender difference in the share of listening going to female artists.Source: Studies of Gender Representation · p. 79
  • Goodreads users were more likely to give high ratings to authors of their own gender, while these studies examine bias sources and propagation through standard recommendation algorithms.The monograph distinguishes these behavioral studies from work focused chiefly on defining provider-fairness metrics and optimization strategies.Source: Studies of Gender Representation · p. 79

Ensuring Fair Representation

  • Greedy fair re-ranking can preserve the order of items within each group while maximizing overall utility subject to a binomial fairness constraint, assuming accurate utility estimates.Empirical studies report no substantial utility loss in their experimental settings; randomization can also prevent the same items from always being displaced while preserving an individual fairness bound.Source: Ensuring Fair Representation · pp. 79–80

6.2 Provider Exposure and Utility

  • Exposure-based constructs treat position-weighted attention as a provider resource and relate provider utility to consumer utility, assuming relevance measures proxy item value.Unlike representation measures that count list appearances, exposure approaches account for predicted user-item match and the greater attention typically given to higher-ranked items.Source: §6.2 · p. 80
  • Because exposure differs by rank and lists offer few opportunities, fairness is often evaluated across sequences or stochastic rankings, but this suits only contexts where aggregate utility over time is appropriate.Recruitment lists are an example where each ranking may still need to be fair with respect to protected groups.Source: §6.2 · pp. 80–81

Individually-Fair Exposure

The section surveys exposure-based fairness for individual items and provider groups, then outlines reranking and learning approaches that seek to improve fair exposure. It also notes that aggregation across requests can obscure whether the right items are exposed for the right needs.

Individually-Fair Exposure

  • Expected exposure compares a ranking policy with a relevance-respecting target to measure whether relevant items receive comparable exposure, with item exposure aggregatable to providers.The section also describes equitable attention, which amortizes provider exposure across rankings in proportion to utility.Source: Individually-Fair Exposure · pp. 81–82
  • Aggregating exposure across requests can detach exposure from the information need, while some metrics also lack interpretable scales and support comparison only on the same request distribution.Aggregating attention and relevance separately can allow a system to expose items on the wrong queries.Source: Individually-Fair Exposure · p. 82

Group-Fair Exposure

  • Group exposure metrics compare groups’ exposure with relevance or a target, but aggregation can reward exposing the wrong items within a group rather than the items that best serve users’ needs.Item-level exposure differences followed by group aggregation are used by the TREC 2022 Fair Ranking track.Source: Group-Fair Exposure · pp. 83–84

Ensuring Fair Exposure

  • Fair-exposure interventions include reranking, constrained stochastic ranking, and learning objectives that limit exposure disparities relative to differences in utility.Other approaches regularize for statistical independence or balance recommendation neighborhoods; cold-start recommendation can maximize the minimum exposure of new items.Source: Ensuring Fair Exposure · pp. 84–85

6.3 Fair Accuracy and Pairwise Fairness

The section surveys pairwise fairness, subject fairness, and fairness over time in information access. It describes dynamic evaluation as simulation of system evolution, while noting that evaluation methods for fairness over time are still emerging.

6.3 Fair Accuracy and Pairwise Fairness

  • Pairwise fairness asks whether relevant items are ordered accurately at similar rates across provider groups, with intra-group and inter-group forms for binary protected groups.Two-stage relevance feedback can help avoid optimizing only for click probabilities; pairwise accuracy can be sampled, though its potential sample-efficiency advantage remains underexplored.Source: §6.3 · pp. 85–86

6.4 Related Problem: Subject Fairness

  • Subject fairness shares provider-fairness metrics and mechanisms but is closely tied to diversity, and may require machine-readable subject labels that often need content analysis or human annotation.The literature has studied image search more than other fair recommendation and retrieval contexts, and the area remains comparatively under-attended.Source: §6.4 · pp. 86–87

Dynamic Fairness

  • Information access changes as users and items arrive or leave and user interests shift, yet fairness research on these dynamics remains limited.The section distinguishes these changing system populations and interests from dynamic-learning settings where the environment’s items and users are treated as static.Source: Dynamic Fairness · p. 88

7.1 Feedback Loops

  • Presentation-driven feedback can magnify small popularity differences into unequal exposure, making it harder for new items to gain traction and requiring fairness-aware recommenders to counter the tendency.Studies report increased recommendation consistency across users alongside increased inequity of item exposure.Source: §7.1 · p. 89

7.2 Dynamic Evaluation

  • Dynamic evaluation simulates user arrival, recommendations, responses, and periodic retraining; fairness can be tracked across intervals and evaluation cycles, but methods remain emerging.Off-policy evaluation is needed when user responses are observed for only a subset of possible recommendations.Source: §7.2 · p. 90

7.3 Opportunities in Feedback Loops

The concluding sections map open problems and offer recommendations for studying fairness in information access. The authors characterize their monograph as a prospective map of a young field and urge contextualized goals, metrics, data, and reporting.

7.3 Opportunities in Feedback Loops

  • Feedback can adjust fairness objectives over time in response to observed deviations, while reranking can account for past rankings to improve aggregate fairness.One approach adjusts objectives for the next recommendation batch; another considers rankings accumulated up to the current time.Source: §7.3 · p. 90

Next Steps for Fair Information Access

  • Information access fairness requires attention to ranked, multisided systems, stakeholder concerns, and data needs, alongside care to avoid overbroad or ungeneralizable claims.The field connects longstanding information-access concerns with algorithmic fairness research.Source: Next Steps for Fair Information Access · p. 91

8.1 Directions for Future Research

  • Open research includes multidimensional and intersectional bias, participatory study of affected stakeholders, dynamic fairness, additional stakeholders, user responses, and better metrics.The agenda also calls for studying bias across applications and improving data and model provenance documentation.Source: §8.1 · pp. 92–93

8.2 Recommendations for Studying Fairness

  • Researchers should define a contextualized social goal, operationalize it with a justified metric, use appropriate data, and carefully report limitations.The monograph cautions against statistical inference techniques for annotating individual people.Source: §8.2 · pp. 94–95

8.3 Concluding Remarks

  • Because the field is too young for a comprehensive retrospective, the monograph assembles an incomplete prospective map that future research can extend.The authors anticipate that further work will fill in details and identify new dimensions.Source: §8.3 · p. 95

Appendix A

The appendix points readers to datasets, software, and other resources for studying fair information access. It notes that open-source software remains limited and that resource links may degrade over time.

Resources for Fair Information Access

  • The appendix collects fair information access resources and cautions that its links may become outdated more quickly than publication references.The pointers are presented as current to the time of publication.Source: Resources for Fair Information Access · p. 125

A.1 Data Sets

  • Available datasets include TREC Fair Ranking data for provider fairness in academic and Wikipedia search, PIReT tools for book recommendation data, and datasets developed for fair ranking.The PIReT tools integrate recommendation datasets with book and author metadata.Source: §A.1 · p. 125

A.2 Sofware

  • Open-source fair recommendation and retrieval software is not yet widely distributed; resources include Terrier’s xQAD, experimental scripts, and librec-auto.librec-auto supports recommender-system experiments, including fairness metrics.Source: §A.2 · pp. 125–126

Related papers

Loading 2105.05779v4…