Fair Feed Ranking for Participatory Budgeting

Carina Hausladen

MünchenBudget

Scale

• Every proposal must be assessed for feasibility, legality, cost, and duplication.
  • Centralized review is
    costly and it
  • concentrates agenda-setting power.

Scale creates two problems

Vetting.

• Every proposal must be assessed for feasibility, legality, cost, and duplication.
  • Centralized review is
    costly and it
  • concentrates agenda-setting power.

• one of the oldest topics in political science
• whoever controls which options reach the vote can often decide the outcome

Why proposals were "not feasible" in 2026

Vetting.

Scale creates two problems

Scale creates two problems

• Citizens cannot inspect thousands of proposals before voting.
• The set of proposals a voter encounters is shaped by
  • the voter’s entry point
  • the platform’s ordering rule

filtering

Popularity signals make the popular more popular

filtering

Cross-Platform-Comparison

Is "random" our best bet?

Recommender systems

Recommender systems

Recommender systems

How could a
fair feed look like?

Ranking process

1. transparent
2. steerable
3. address the vetting problem
4. exposure correction / visibility floor

Simulation

Results

Results

FairFeed Prototype

Text

Text

• Cold start, language-based preferences (Sanner et al. 2023)
• Graded score (Zhuang et al. 2024)
  • \( \displaystyle s_i = \mathrm{LLM}(\text{brief},\ \text{proposal } i) \in [0,1] \)
• Beta prior (Austin et al. 2024)
  • \( \displaystyle f(x) \propto x^{\alpha_i - 1}\,(1 - x)^{\beta_i - 1} \)
  • \( \displaystyle \alpha_i = \kappa\,s_i \qquad \beta_i = \kappa\,(1 - s_i) \qquad \kappa = 4 \)
  • \( \displaystyle \mu_i = \frac{\alpha_i}{\alpha_i + \beta_i} \qquad \sigma_i = \sqrt{\frac{\alpha_i\,\beta_i}{(\alpha_i + \beta_i)^2\,(\alpha_i + \beta_i + 1)}} \)

Text

• Pick the four closest to the brief
  • \( \displaystyle \text{next card} = \arg\max_{i \notin A} \ \left( \mu_i + c\,\sigma_i \right) \)
• Bayesian preference elicitation (Austin et al. 2024)
  • rated card j
    • \( \displaystyle u_{j} \in \{0,\ 0.5,\ 1\} \)
    • \( \displaystyle e_{ij} = (2u_{j} - 1)\,r_{ij} \)
    • \( \displaystyle \alpha_i \leftarrow \alpha_i + \max(e_{ij},\,0) \)
    • \( \displaystyle \beta_i \leftarrow \beta_i + \max(-e_{ij},\,0) \)
  • Each rating updates the posterior of every proposal

Text

• Exploration–exploitation trade-off (Chen et al. 2019; Gupta et al. 2021)

  • Explore slots:

    • underexplored proposals

  • Exploit slots:

    • \( \displaystyle \text{slots } 1 \ldots K' = \text{the } K' \text{ largest } \mu_i  \)

Open Questions

Conclusion

• PB cycles run at large scales

• Platform "solutions" partly aggravate the problem

• FairFeed combines
  • fair exposure
  • crowd-sourced-vetting through negative feedback
  • transparent ranking on explicit preferences

• FairFeed increases
  • browsing length and relevance
  • coverage
  • non-seed votes
  • and top 20 quality

Appendix

Madrid city open data (CC BY 4.0)
 

• 132 proposals from 2021/22
  (translated from Spanish into English)
• 8 categories:
  • Pedestrian streets & public space
  • Culture, community & heritage
  • Parks, trees & green space
  • Traffic safety & lighting
  • Pavements & accessibility
  • Cycling
  • Play areas & children
  • Cleanliness, waste & environment

Text

Other onboarding ideas?

Can we implement more "transparency"?

Is recommending interesting items enough?

• Cold start (Sanner et al. 2023)
    No interaction history for this person. The ranking starts from text alone.

•   Language-based preferences (Sanner et al. 2023)
    Preferences given as free text, not as ratings, clicks or likes. No collaborative filtering.

•   Bayesian preference elicitation (Austin et al. 2024)
    The brief sets the prior; each rating and vote updates the posterior of every proposal, before the feed and inside it.

• Exploration–exploitation trade-off (Chen et al. 2019; Gupta et al. 2021)
    A multi-armed bandit over a fixed pool: most slots rank by the current estimate, a few are sampled from the rest.

• Read all 1,043 titles, in three passes.
• close candidates picked
• 44 close calls re-read with the description capped at 160 characters
• Score as 
  • A — redundant duplicate: same thing, no location. Funding both is waste.
  • B — positional variant: same thing, different site, same €100k.
  • C — same goal, different means.

• A user who lacks domain knowledge cannot state a preference over a feature until the system has shown them  some.
• Interviewing the participant over the pool itself:
  • showing a fixed set of (unrated) proposals to the user
  • the system chooses the one whose rating would teach it the most about the user 
• Example:
  • Eight turns, each turn shows one proposal, the participant rates it,
  • the scores over the pool are updated,
  • and the next proposal is picked against the updated scores.

How to improve the onboarding?

Mechanisms for identifying weak, infeasible, duplicate, or low-quality content.

Learning from online communities

Learning from online communities

• Common characteristic: large option space
• Found ways to manage them effectively

Transparency

•  The LLM rewrites the profile sentence using the 5 ratings.
•   The feed is ranked on that new sentence.

Calibration

• The app shows one proposal, and the app rates it
• The guess is updated for every proposal
• The next proposal is the one the app is least sure about.
• This repeats 5 times.

Language brief

• citizens write in their own words what you care about
• The LLM turns that into a short profile sentence.
• The LLM then gives every proposal a score from 0 to 1 for how well it fits the brief.

Three steps

Should we keep categories?

• Getting citizens interested
  • beyond their own needs / their initial favorite
  • towards their communities' needs

Zielnicki et al. 2026 Netflix viewing logs, ~2M users

- recommendation triples its chance of being watched

- 12–16% is the power of personalization.

 - Most-popular rows: −12%

 - Random rows: −16%

Learning from recommender systems

1. intent fidelity

2. ease of use

3. procedural transparency

Stages

1. Planning: LLM input

2. Sourcing: Finding candidate proposals

3. Curation: scoring proposals

4. Ranking

Attention

• Citizens cannot realistically inspect thousands of proposals before voting.
• The set of proposals a voter encounters is shaped by
  • the voter’s entry point
  • the platform’s ordering rule
• Two margins
  • Extensive margin: does a voter encounter proposals beyond the one that initially brought them to the platform?
  • Intensive margin: is attention concentrated on already-visible proposals, or do less prominent but potentially valuable proposals receive meaningful exposure?
• A democratic process is not judged only by how many people can participate, but by what kind of participation it enables.
  • If PB platforms fail on either margin, they deny citizens a meaningful chance to encounter alternatives and form considered judgments [9, 14].
  • Fair exposure is therefore best understood as a floor on visibility: without it, a proposal’s failure may reflect rejection, but it may equally reflect lack of exposure.

Scale creates two problems

Vetting.

• Every proposal must be assessed for feasibility, legality, cost, and duplication.
• Centralized review is not only costly, but it also concentrates agenda-setting power.
• Delegating evaluation back to citizens avoids this concentration of power, yet reintroduces the attention problem.

Scale creates two problems

PB at scale

Mechanisms for identifying weak, infeasible, duplicate, or low-quality content.

Learning from online communities

sorting

filtering

• used by more than 200 public institutions
• in over 35 countries
• and serves more than 100 million people.

Cross-Platform-Comparison

Munich field data, June 2026