Research map / Label-efficient and robust learning / Recognition and long-tail
Representative Ranking for Deliberation in the Public Sphere
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces justified representation (JR) as a constraint on comment ranking so that platform-optimized sets (engagement, diverse approval, Perspective API bridging scores) still represent all large cohesive user groups, proving low price of JR under clustered preferences and validating with GreedyCC on Remesh campus-protest data.
Problem
Algorithmic comment ranking that optimizes civility or bridging or engagement can suppress legitimate viewpoints; deliberation requires representation of diverse views, which existing ranking methods do not guarantee.
Approach
Adds the social-choice axiom of justified representation as a hard constraint on the top-k comment set, maximizing any score function subject to JR, and analyzes the price of JR across score-function classes and under Mallows mixture preference models.
Key results
- Enforcing JR with GreedyCC reduced unrepresented users from 18% to 5% (engagement), 15% to 4% (diverse approval), and 4% to 2% (Perspective API).
- Average price of GreedyCC over ten Remesh questions (k=8) was 1.05 for engagement, 1.06 for maximin diverse approval, and 1.18 for Perspective API bridging.
- 47 of 48 JR feeds also satisfied EJR+; Perspective API rankings satisfied JR 100% of the time without enforcement.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Enforce JR with GreedyCC | Engagement ranking · Remesh campus protest discussions, 10 questions | unrepresented users: 18% -> 5% | Helped |
| Enforce JR with GreedyCC | Maximin diverse approval ranking · Remesh campus protest discussions, 10 questions | unrepresented users: 15% -> 4% | Helped |
| Enforce JR with GreedyCC | Perspective API ranking · Remesh campus protest discussions, 10 questions | unrepresented users: 4% -> 2% | Helped |
| Enforce JR with GreedyCC | Engagement ranking · Remesh campus protest discussions, 10 questions | engagement score: 1.05 | Hurt |
| Enforce JR with GreedyCC | Maximin diverse approval ranking · Remesh campus protest discussions, 10 questions | maximin diverse approval score: 1.06 | Hurt |
| Enforce JR with GreedyCC | Perspective API ranking · Remesh campus protest discussions, 10 questions | Perspective API bridging score: 1.18 | Hurt |
| Use GreedyCC to enforce JR | Comment ranking across scoring functions · Remesh campus protest discussions | representation of political groups: 100%, 100%, and 70% | Helped |
Surprising findings
- Perspective API bridging-attribute ranking satisfied JR in all instances without explicit enforcement, more often than engagement (30%) or diverse approval (50%).
- Maximin diverse approval can have a worst-case price of JR equal to k, meaning optimal JR sets may attain only 1/k of the unconstrained diverse approval.
Limitations
- Approval matrix is inferred from sparse votes in Remesh data rather than observed, and inference bias could affect results.
- Experiments are offline re-rankings of historical data; no A/B test on live users.
- Open question of which users belong in the approval matrix for a given platform.
Experimental setup
Datasets: Remesh campus protests dataset (approx. 300 participants per session, two sessions, ten questions)
Benchmarks: Remesh campus protests dataset
Metrics: user representation rate JR satisfaction rate price of GreedyCC diverse approval engagement Perspective API bridging score
Baselines: engagement ranking maximin diverse approval ranking Perspective API bridging ranking
Builds on: justified representation (Aziz et al.) GreedyCC Mallows model Perspective API
Closest papers by meaning
- DEBATE: A Large-Scale Benchmark for Evaluating Opinion Dynamics in Role-Playing LLM Agents (NeurIPS 2026 · similarity 0.84)
- Examining Alignment of Large Language Models through Representative Heuristics: the case of political stereotypes (ICLR 2025 · similarity 0.84)
- Generative Social Choice: The Next Generation (ICML 2025 · similarity 0.84)
- Debating with More Persuasive LLMs Leads to More Truthful Answers (ICML 2024 · similarity 0.83)
- Political Neutrality as Balanced Approval: A Large-Scale Human Evaluation of AI Responses (NeurIPS 2026 · similarity 0.83)
- Embeddings for Preferences, Not Semantics (NeurIPS 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.