🤖 AI

Algorithmic Approaches to Opinion Selection for Online Deliberation: A Comparative Study

This paper benchmarks existing algorithmic strategies for selecting opinions in online deliberation and proposes a novel social-choice-inspired algorithm that empirically achieves the strongest trade-off between proportional representation and diversity, addressing the risk of algorithmic automation erasing minority voices.

Original authors: Salim Hafid, Manon Berriche, Jean-Philippe Cointet

Published 2026-02-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Salim Hafid, Manon Berriche, Jean-Philippe Cointet

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are hosting a massive town hall meeting with 1,000 people. Everyone has an opinion, a story, or a suggestion to share. By the end of the night, you have a mountain of 5,000 sticky notes on the wall.

Now, imagine you need to present the "Top 5" most important ideas to the Mayor, who only has 5 minutes to listen.

The Problem:
If you just pick the 5 most popular notes (the ones with the most sticky dots), you might miss out on the quiet but crucial voices of a small group. If you try to pick 5 totally different notes just to be "diverse," you might end up ignoring the 60% of people who all agreed on one big issue. If you try to find the "middle ground" to make everyone happy, you might end up with 5 boring, watered-down ideas that nobody really cares about.

This is the challenge of Online Deliberation. How do you algorithmically pick a small, fair, and representative summary of a huge crowd's thoughts?

The Paper's Solution:
The authors of this paper are like "Digital Matchmakers." They tested different computer programs (algorithms) to see which one does the best job of picking those "Top 5" notes without leaving anyone out or repeating the same idea twice.

Here is a breakdown of their findings using simple analogies:

1. The Old Ways (The Baselines)

The researchers tested three common ways computers usually pick opinions:

The "Popularity Contest" (Engagement):
- How it works: The computer picks the 5 notes with the most votes.
- The Flaw: It's like a reality TV show. The loudest, most popular voices win, and the minority groups (the "niche" ideas) get completely erased. It's efficient, but not fair.
The "Bridge Builder" (Bridging):
- How it works: The computer looks for ideas that everyone agrees on, even if they are usually enemies. It tries to find common ground.
- The Flaw: It's great for peace, but it often ignores the "disruptive" or "radical" ideas that are necessary for a healthy debate. It flattens the conversation to avoid conflict.
The "Mixer" (Diversity):
- How it works: The computer tries to pick 5 notes that are as different from each other as possible.
- The Flaw: It might pick 5 weird, unpopular ideas just to be different, leaving the majority of the crowd feeling unrepresented.

2. The New Hero: "DiverseBJR"

The authors created a new algorithm called DiverseBJR. Think of this as a Fairness-First Chef.

The Recipe:
1. The "Fair Share" Rule (BJR): First, the algorithm ensures that if a group of people makes up 20% of the crowd, they must get roughly 20% of the "Top 5" spots. It's like cutting a pizza so every slice of the pie gets a fair piece of the toppings, not just the people who shouted the loudest.
2. The "No Duplicates" Rule (Diversity): Once the fair shares are assigned, the algorithm checks: "Are these 5 ideas actually different?" If two ideas are almost the same (like two notes saying "We need more parks"), it swaps one out for a totally different idea (like "We need better schools") to make sure the final list covers the whole spectrum of thought.

3. The Results: Why It Matters

The researchers ran this on real data from political debates about protests. Here is what they found:

Small Groups Matter Most: When you only have room for a very small number of opinions (like 2 or 3), the new algorithm is a superhero. It manages to include the minority voices and keep the list diverse, without needing to sacrifice the majority's voice.
The "Sweet Spot": The old "Popularity" method fails at fairness. The "Bridge" method fails at diversity. The new DiverseBJR method finds the "Goldilocks zone." It gives you a list that is small enough to be readable, but fair enough that no one feels silenced.

The Big Picture Metaphor

Imagine you are curating a museum exhibit about a war.

The Popularity Algorithm would only show the generals' speeches.
The Diversity Algorithm would show 5 random, unrelated artifacts that tell no coherent story.
The Bridge Algorithm would only show the peace treaties, ignoring the battles.
The DiverseBJR Algorithm ensures you have a soldier's letter, a civilian's diary, a general's map, and a peace activist's poster. It guarantees that every major "faction" of the story gets a spot on the wall, but it also makes sure the exhibits aren't just 5 copies of the same soldier's letter.

In Conclusion:
This paper proves that we don't have to choose between "fairness" and "diversity." By using a little bit of math from social choice theory (the study of voting), we can build computer systems that automatically select the best, fairest, and most interesting summaries of our collective conversations. This is crucial for the future of online democracy, ensuring that when we vote or debate online, the computer doesn't accidentally silence the quiet voices.

1. Problem Statement

In online deliberation platforms (e.g., Pol.is, Remesh), participants generate large volumes of free-form opinions, critiques, and proposals. Human facilitators cannot process all inputs to create digestible reports for stakeholders. Consequently, algorithmic selection is used to choose a small, representative subset of opinions ( $k$ ) from a larger set ( $m$ ).

The core challenge is that existing selection strategies often prioritize specific, sometimes conflicting, democratic goals:

Consensus/Popularity: Selecting the most popular opinions (e.g., Engagement, Bridging) often ignores minority voices and reduces content diversity.
Diversity: Selecting diverse opinions often fails to ensure proportional representation, meaning large groups of voters may have no opinion in the final subset that represents their views.
The Gap: There is a lack of systematic comparison of these strategies against democratic desiderata, and no existing algorithm effectively balances proportional representation (ensuring minority groups are represented) with diversity (ensuring the selected opinions are distinct and non-redundant).

2. Methodology

Theoretical Framework

The authors frame the problem as an Approval-Based Committee Election.

Inputs: A set of $n$ users and $m$ opinions. Users provide a binary approval profile ( $A_u \subseteq [m]$ ).
Goal: Select a subset $S \subseteq [m]$ of size $k$ .
Axioms Used: The study leverages Social Choice Theory, specifically:
- Justified Representation (JR): Ensures any cohesive group of size $\ge n/k$ has at least one approved opinion in $S$ .
- Balanced Justified Representation (BJR): A stronger variant requiring that the selected opinions are assigned to users in a balanced way (each opinion represents $\approx n/k$ users), preventing a few popular items from dominating the representation.

Proposed Algorithm: DiverseBJR

The authors propose DiverseBJR, a novel selection rule that integrates BJR with diversity constraints.

Mechanism: It uses a two-stage greedy approach (similar to GreedyCC):
1. Stage 1 (Satisfying BJR): Iteratively selects opinions to satisfy "BJR budgets" (ensuring each selected opinion represents a specific quota of users).
2. Stage 2 (Filling Slots): Fills remaining slots if Stage 1 cannot satisfy the quota.
Diversity Integration:
- Tie-breaking: When multiple opinions satisfy the current BJR budget, the algorithm breaks ties by favoring opinions with fewer "ε-neighbors" (opinions that are too similar based on Hamming distance of approval vectors).
- Ineligibility: After selecting an opinion, its ε-neighbors are marked ineligible for future selection only if a feasibility check (a greedy simulator) confirms that the remaining BJR constraints can still be satisfied.
Distance Metric: Uses Normalized Hamming Distance on approval vectors. Unlike Jaccard similarity, Hamming distance treats shared "disapprovals" (zeros) as informative, capturing both support and opposition patterns.

Evaluation Metrics

The study evaluates algorithms against five democratic criteria:

Unrepresentation (Individual & Group): Percentage of users/groups with no approved opinion in $S$ .
Consensus: Maximin cross-group agreement (does the set contain an opinion approved by all groups?).
Coverage Gap (CG): The maximum distance between any non-selected opinion and its closest selected opinion (measures how well $S$ covers the opinion space).
Redundancy: The fraction of selected opinions that are "clones" (within distance $\epsilon$ of another selected opinion).
Social Welfare: Total number of approvals covered.

Experimental Setup

Datasets:
1. Remesh "Polarized Issues Dataset": Real-world US public dialogues on protest rights (N $\approx$ 300). Includes inferred approval matrices. Split into "Controversial" and "Consensual" questions.
2. Virtual Citizen Assembly (UK): A dataset with fully observed votes (no inference needed) to validate robustness.
Baselines:
- Consensus-based: Engagement (popularity), Bridging (maximin cross-group agreement).
- Diversity-based: Greedy minimization of Coverage Gap.
- Representation-constrained: Approximations of JR and BJR.
- Ours: DiverseBJR.
- Control: Random selection and Zero-shot ChatGPT (excluded from main results due to cardinality issues).

3. Key Contributions

Systematic Benchmarking: The first comprehensive comparison of algorithmic opinion selection strategies (Consensus, Diversity, JR, BJR) using metrics tailored to democratic deliberation.
Novel Algorithm (DiverseBJR): An algorithm that operationalizes Social Choice axioms (BJR) while explicitly optimizing for diversity without violating proportional representation guarantees.
Empirical Evidence: Demonstrates that while no single strategy dominates all metrics, DiverseBJR offers the optimal trade-off for small subset sizes ( $k \approx 1-5$ ), which is typical for deliberation reports.
Open Source: Release of code and evaluation protocols to facilitate future research and practitioner adoption.

4. Results

The experiments focused on controversial questions where selection is most difficult:

Representation: Representation-constrained algorithms (JR, BJR, DiverseBJR) significantly outperform consensus-based methods (Engagement, Bridging) in reducing unrepresented users, especially for small $k$ . DiverseBJR achieves the strongest early gains.
Coverage & Diversity:
- Pure Diversity algorithms achieve the lowest Coverage Gap (best space coverage) but fail to guarantee proportional representation.
- DiverseBJR achieves a Coverage Gap close to pure Diversity while maintaining the high representation guarantees of BJR.
Redundancy:
- Consensus-based methods (Engagement) select highly clustered, redundant opinions.
- DiverseBJR significantly reduces redundancy compared to BJR and Engagement, spreading selections across the opinion space.
Consensus: DiverseBJR maintains consensus levels similar to Engagement and BJR, proving that enforcing proportionality and diversity does not necessarily sacrifice broad acceptability.
Trade-off: DiverseBJR provides the best Pareto frontier for small $k$ : it improves proportional representation relative to diversity baselines while preserving high coverage and low redundancy.

5. Significance and Impact

Democratic Integrity: The study highlights that automated selection is not neutral; it shapes deliberation outcomes. Using algorithms that ignore minority representation (like pure popularity) can erode democratic legitimacy.
Practical Application: For platforms like Remesh or Pol.is, DiverseBJR offers a deployable solution to generate "digestible" reports that are both representative (fair to all groups) and diverse (covering the full spectrum of views without redundancy).
Theoretical Advancement: It bridges the gap between Social Choice Theory (which offers fairness guarantees) and Recommender Systems (which optimize for diversity), creating a hybrid framework suitable for high-stakes democratic contexts.
Future Directions: The authors suggest deploying these algorithms in live deliberation settings to observe real-world behavioral impacts, moving beyond offline simulation.

In summary, the paper argues that for online deliberation to be truly democratic, selection algorithms must move beyond simple popularity or diversity heuristics. Instead, they should adopt Social Choice-inspired rules like DiverseBJR that mathematically guarantee proportional representation while actively promoting the diversity of the selected opinion set.