Carina Ines Hausladen PRO
I am an Assistant Professor for Computational Social Science at the University of Konstanz, Germany.
Fair Feed Ranking for Participatory Budgeting
Carina Hausladen
1 Million Euro
MünchenBudget
Scale
Vetting.
- Every proposal must be assessed for feasibility, legality, cost, and duplication.
- Centralized review is
costly and it - concentrates agenda-setting power.
- Centralized review is
Scale creates two problems
1059 reviewed
461 "feasible"
Vetting.
- Every proposal must be assessed for feasibility, legality, cost, and duplication.
- Centralized review is
costly and it - concentrates agenda-setting power.
- Centralized review is
- 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
Crowd-sourced
Vetting.
Scale creates two problems
Attention.
- 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
Scale creates two problems
Attention.
- 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
sorting
filtering
2025
31%
2026
18%
2025
13%
2026
38%
Share of proposals
with at least one comment
Share of comments written
by the proposal's own author
Vote accumulation
Daily correlation
with final ranking
- 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
the most popular petitions gained "at the expense of those with fewer signatories."
Cross-Platform-Comparison
- Shuffling is on by default, for every voter.
- Each time a voter opens the ballot, the order is drawn fresh.
- Shuffling is on by default, for every voter.
- Each time a voter opens the ballot, the order is drawn fresh.
Is "random" our best bet?
| Company | Paper | Result for random ordering |
|---|---|---|
| Netflix | Zielnicki et al. 2026 | −16% engagement |
| Wang et al. 2018 Qin et al. 2020 | −14% in email search, −31% in file-storage search. “Unavoidably degrades the user experience”, even when only “swapping adjacent pairs of items”. | |
| Liu et al. 2017 | “unranked results are lower quality” | |
| Duolingo | Yancey & Settles 2020 | −0.5% daily active users, −2% new-user retention. |
| Relevance | |
| Breadth | |
| Depth | |
| Exposure |
| Support citizens in investing more time in desirable tasks. | |
| Move citizens from self-centred needs to also consider community needs. | |
| Better decisions. Citizens compare alternatives before they choose. | |
| Every submitted proposal should have a fair chance of being seen. |
Recommender systems
| Commercial goal | Democratic goal | |
|---|---|---|
| Relevance | ||
| Breadth | ||
| Depth | ||
| Exposure fairness |
| Maximise user engagement. Time spent is revenue. | Support citizens in investing more time in desirable tasks. Taking part is part of being human. | |
| Diversify topics to prevent user boredom. | Move citizens from their own needs to the community’s needs. | |
| Retain sellers and creators on the platform. | No window dressing. Letting 14-year-olds submit proposals means little if nobody ever sees them. |
Recommender systems
Trinkwasserbrunnen
2 winners · Trinkbrunnen · Wasserspender · Kostenlose Spender
~30
Größere Mülleimer
Mülleimer mit Deckeln · Pfandring · Papierkörbe
~25
Bäume pflanzen
Einfach ein Baum · Schirmbäume · Obstbäume
~12
Wildblumenwiesen
2 winners · Wildblumen · Blütenwiese · Streuobstwiese
~10
Wasserbänke
Wassersprühanlagen · Sprühnebel · Sonnensegel
~7
TanzPark-ett
Tanzlinde · Tanzfläche mit Spiegelwand · Holzboden
3
Glatte Bordsteinkanten
Schwellen und Kanten beseitigen
1
Defi im öffentlichen Raum
none
0
MünchenBudget 2025 winners
| Commercial goal | Democratic goal | |
|---|---|---|
| Relevance | ||
| Breadth | ||
| Depth | ||
| Exposure fairness |
| Maximise user engagement. Time spent is revenue. | Support citizens in investing more time in desirable tasks. Taking part is part of being human. | |
| Recommend similar items to keep users scrolling. | Better decisions. Citizens move beyond their uninformed first impression. | |
| Retain sellers and creators on the platform. | No window dressing. Letting 14-year-olds submit proposals means little if nobody ever sees them. |
Recommender systems
of 1,043
reaches ballot
winners
win rate
Gesundheit
23%
56%
5
3.8%
Umwelt & Nachhaltigkeit
41%
49%
7
3.3%
Bildung & Kultur
18%
47%
2
2.3%
Stadtgestaltung
45%
45%
4
1.9%
Freizeit & Sport
32%
50%
3
1.8%
Kinder, Jugend & Familie
33%
47%
2
1.2%
Soziales & Miteinander
49%
49%
3
1.2%
Mobilität
29%
36%
1
0.9%
Sonstiges
9%
36%
0
0.0%
Digitales
6%
30%
0
0.0%
Bars: share of submissions passing the eligibility screen. Multi-valued labels, so column 1 sums past 100%.
MünchenBudget 2025 winners
| Relevance | |
| Breadth | |
| Depth | |
| Exposure |
| Support citizens in investing more time in desirable tasks. | |
| Move citizens from self-centred needs to also consider community needs. | |
| Better decisions. Citizens compare alternatives before they choose. | |
Votes per proposal in 2026
Top 46 proposals
38% of all votes
| Relevance | |
| Breadth | |
| Depth | |
| Exposure |
| Support citizens in investing more time in desirable tasks. | |
| Move citizens from self-centred needs to also consider community needs. | |
| Better decisions. Citizens compare alternatives before they choose. | |
How could a
fair feed look like?
Ranking process
- transparent
- steerable
- address the vetting problem
- exposure correction / visibility floor
- München-Budget 2025
- 1,043 proposals submitted;
- 459 feasible
- 10 funded
- 4 vetoed
- 7134 voters
- 640 comments
- present or not
- pseudonymised author
- present or not
- 1,043 proposals submitted;
Simulation
Modeling choices
- decision
- seed is supported automatically
- each other proposal is supported as a combination of fit, divisiveness, position
- browsing length
- rises with recent relevance and
- falls with fatigue
Relevant vs browsed
Exposure coverage
Results
Non-seed votes
Top-20 quality
Results
FairFeed Prototype
Text
Wang et al. (2024)
- \( \mathrm{X}_{i} \) = Sentence-BERT( text of proposal i )
- \( \displaystyle \mathrm{sim}_{ij} = \exp\left(-\tfrac{\lambda}{2}\,\lVert X_i - X_j \rVert^2\right) \)
- similarity, labelled by an LLM:
- \( \displaystyle v_{ij} \in \{0,\ 0.5,\ 1\} \)
- \( \displaystyle r_{ij} = (2v_{ij} - 1)\,\mathrm{sim}_{ij} \)
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
- rated card j
Text
- Exploration–exploitation trade-off
(Chen et al. 2019; Gupta et al. 2021)-
Explore slots:
-
underexplored proposals
-
-
Exploit slots:
-
\( \displaystyle \text{slots } 1 \ldots K' = \text{the } K' \text{ largest } \mu_i \)
-
-
| E | E | E | E | x | E | E | E | E | x | E | E | E | E | x | E | E | E | E | x | … |
| ↑ | ↑ | ↑ | ↑ |
E exploit x explore
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
FairFeed prototype
- Language-based preferences
- Bayesian preference updating
Can you help us to improve our prototype?
- Pseudonymized vote records
- vote count per user
- # proposals viewed
- Proposal texts licensed under CC BY 4.0
carina.hausladen@uni-konstanz.de
ConsulCon 2026
By Carina Ines Hausladen
