PaperlayerView exact revision on arXiv

Source-linked AI summary

Fair Feed Ranking for Participatory Budgeting

Carina I. Hausladen

arXiv:2609.29819v1cs.CYcs.IR

TL;DR

In large-scale participatory budgeting, citizens cannot inspect the full proposal pool, making feed order a form of agenda-setting power. The paper proposes and simulates FairFeed, which ranks on declared preferences and corrects exposure; the simulation reports broader discovery, more even visibility, greater cross-cutting support, and better resistance to manipulation. The authors identify human-subjects evaluation of preference elicitation as a next step.

  • Problem

    In large-scale participatory budgeting, citizens cannot inspect all proposals, and the ordering rule shapes which alternatives they encounter and how much visibility proposals receive.

    Source: §1 · p. 1

  • Method

    FairFeed ranks using declared preferences, boosts underexposed proposals, and includes a rate-limited reject channel, with its ranking logic evaluated in a simulation anchored in MünchenBudget data.

    Source: §1 · p. 2, §2.1 · p. 3

  • Results

    In the MünchenBudget-anchored model, FairFeed broadens discovery, spreads exposure across the eligible pool, roughly doubles cross-cutting support, and improves shortlist quality and resistance to manipulation.

    Source: §4 · p. 7

  • Takeaways & Limitations

    Making the hidden consequences of ranking rules visible before deployment is a useful first step because trust in civic procedures is hard to build and easy to lose.

    Source: §4 · p. 7

    The simulation proxies elicited preferences with broad topic categories, so human-subjects evaluation must test whether onboarding recovers real interests accurately enough for deployment.

    Source: §4 · p. 7

Abstract

from arXiv · show

In large-scale participatory budgeting, citizens cannot inspect the full proposal pool, so the order in which proposals are shown becomes a form of agenda-setting power. We argue that fair exposure should therefore be treated as a democratic-design goal. We study Consul Democracy, a widely deployed open-source digital-democracy platform, and show that its proposal feeds are typically ordered by popularity, recency, or comment activity. Building on this diagnosis, we propose FairFeed, a feed-ranking design for PB that uses transparently declared preferences, boosts under-exposed proposals, and admits a rate-limited reject channel for crowd-sourced vetting. We evaluate the design in a simulation anchored in Munich's 2025 PB process and compare it with random, newest, and most-commented feeds. In this simulation, FairFeed broadens proposal discovery, distributes visibility more evenly across the eligible pool, increases cross-cutting support, and improves resistance to manipulation relative to comment-based ranking. We conclude by outlining the human-subjects evaluation needed to test whether onboarding can recover voter preferences accurately enough for deployment in practice.

1 Introduction

Large proposal pools make it difficult for citizens to inspect alternatives and leave both visibility and vetting shaped by platform ordering. The paper diagnoses popularity-, comment-, and recency-based feeds as agenda-setting mechanisms and proposes FairFeed as a fairer design.

1 Introduction

  • Citizens cannot realistically inspect thousands of proposals, while administrators face costly screening and citizen-facing lists still run into the hundreds.The section frames fair exposure as a visibility floor: a proposal may fail because it was rejected or because it was never seen.Source: §1 · p. 1
  • Consul and city deployments rely on popularity, comment-count, and recency orderings that amplify already-visible proposals.The platform’s stock options include most-voted and random, while deployments can offer additional sorts.Source: §1 · p. 2
  • Recommender systems can broaden what residents encounter beyond their own neighborhood, while online-community downvotes offer a precedent for crowd-sourced proposal vetting.The paper notes that negative feedback is typically used sparingly and that PB methods have also used limited negative votes.Source: §1 · p. 2
  • Fair exposure addresses the intensive-margin problem by distributing visibility across items rather than allowing early popularity to determine who is seen.The section also distinguishes discovery beyond a voter’s entry point from the distribution of attention among proposals encountered.Source: §1 · pp. 1–2
  • The paper audits Consul and MünchenBudget, then proposes FairFeed with declared preferences, exposure correction, and a rate-limited reject channel.Its stated aims include broader discovery, more equal visibility, and resilience to manipulation such as coordinated commenting.Source: §1 · pp. 2–3

2 Methods

FairFeed combines declared preferences, exposure correction, and guarded negative feedback in a ranking design evaluated through a simulation calibrated to MünchenBudget. The methods use real process data alongside modeled voters and proposals, with feed differences intended to isolate ranking effects.

2.1 FairFeed

  • FairFeed ranks proposals using declared signals, boosts underexposed items, and permits rejects only after a citizen has supported two other proposals.The onboarding prototype lets citizens state interests in natural language, review and refine mapped want/avoid signals, and receive a personalized feed; preference elicitation itself awaits human-subjects evaluation.Source: §2.1 · p. 3, Figure 1 · p. 3

2.2 Data

  • The data include MünchenBudget’s completed 2025 round of 1,043 proposals and ongoing 2026 round of 1,059 proposals, with proposal and comment authors pseudonymized.In 2025, 459 proposals passed admissibility and the 20 most-liked advanced; 69% of proposals received no comments, while one account produced around 40% of 2026 comments.Source: §2.2 · p. 3

2.3 Modeling PB

  • The simulation models 459 proposals and 7,134 citizens, calibrating 14 proposals with recorded outcomes and drawing the remaining proposals’ qualities from Beta(2, 5).Synthetic proposal quality and reach are assumed independent; a sensitivity check imposes positive correlations and the authors call the independence assumption conservative for the shortlist claim.Source: §2.3 · pp. 3–4
  • Voter support depends on fit-weighted quality relative to a personal threshold, with divisiveness and rank position reducing support probability.The dials for fit slope, divisiveness discount, and position decay are fixed across feeds so differences arise from the feed rather than voters.Source: §2.3 · p. 4
  • Browsing length can respond to recent relevance and fatigue, but discovery and exposure results allow browsing to vary while support, cross-cutting-vote, and quality results hold attention fixed.The browsing coefficients come from literature rather than MünchenBudget fitting, so the authors interpret that browsing rule for direction rather than magnitude.Source: §2.3 · p. 4
  • The model treats any support, neutral rating, or reject as exposure, while support alone accumulates toward the shortlist.Exposure coverage is the share of proposals meeting a visibility floor; the shortlist consists of at most 20 proposals reaching the decisive vote.Source: §2.3 · p. 4
  • Most-commented ranks proposals by a running comment tally, whereas FairFeed uses declared quantities and never reads engagement volume.FairFeed adds a relative underexposure boost below half the average exposure and subtracts normalized accumulated rejects; the comment sort’s feedback loop can amplify visibility and manufactured comments.Source: §2.3 · p. 5

3 Results

In simulation, FairFeed broadens proposal discovery and distributes exposure more evenly, while generating more cross-cutting support and improving shortlist quality with gated rejects. Results also depend on reject rates and broad topic categories, and the simulated fit test cannot fully isolate preference from topical match.

3 Results

  • FairFeed increases exposure to approximately 5.2 proposals per citizen while maintaining near-perfect topical relevance.In the fixed-budget support analysis, 99.5% of browsed proposals are on-topic; the authors note that finer-grained interests could reinforce within-topic bubbles.Source: §3 · p. 5
  • All proposals clear FairFeed’s ≈49-view floor, but only 39% reach ≈97 views, versus 5% under random.Newest and most-commented leave the median proposal seen only nine times; just 7–8% clear the floor under popularity sorts.Source: §3 · p. 5
  • +18%: At an optional-reject rate of ≈0.10, top-20 quality rises from 0.41 to 0.48 over FairFeed’s approve-only ballot.The reject unlocks only after two approvals, and ballot ranking uses supports minus rejects.Source: Figure 2 · p. 6, §3 · p. 6
  • Weaponized rejects lower attack cost from 36 to 25 verified identities, while manufactured comments can cut most-commented top-20 quality from 0.35 to 0.27.Rejects scale with distinct verified identities; one account can post comments without limit.Source: Figure 2 · p. 6, §3 · p. 6
  • FairFeed yields 1.4 cross-cutting votes per citizen on average versus about 0.7 under other feeds; 35% clear the fit bar versus 12% under most-commented.The fit test indicates placement-robust support but cannot fully separate genuine preference from topical match.Source: §3 · pp. 5–6, Figure 2 · p. 6
  • At a reject rate of ≈0.30, quality lifts +32% and divisiveness falls −35%, after which both flatten.The decline in divisiveness at the lower reject rate is partly mechanical; whether rejects remove weak proposals or suppress legitimate contestation requires human-subjects study.Source: §3 · p. 6

4 Conclusion and Outlook

The conclusion frames feed ranking in large-scale participatory budgeting as a democratic-design problem and presents FairFeed as a design addressing attention and vetting. Its MünchenBudget-anchored simulation reports broader discovery, more evenly distributed exposure, and increased cross-cutting support, while real-preference onboarding remains to be tested.

4 Conclusion and Outlook

  • FairFeed addresses large-scale PB’s attention and vetting problems with declared preferences, boosts for under-exposed proposals, and a rationed reject channel for crowd-sourced vetting.The paper frames feed ranking as a democratic-design question because ranking shapes what citizens see while centralized vetting concentrates agenda-setting power.Source: §4 · pp. 6–7
  • FairFeed’s simulation broadens proposal discovery, spreads exposure across the eligible pool, roughly doubles cross-cutting support, and improves shortlist quality and resistance to manipulation.The gated reject channel is materially harder to game than Munich’s current most-commented sort.Source: §4 · p. 7
  • Making ranking consequences visible before deployment is a useful first step because trust in civic procedures is hard to build and easy to lose.Source: §4 · p. 7
  • The simulation uses broad topic categories as proxies for elicited preferences, leaving onboarding accuracy and performance with narrower, fragmented real interests for human-subjects evaluation.Deployment also raises institutional questions about contestability, auditing, and identity verification for the reject channel.Source: §4 · p. 7
Loading 2609.29819v1…