Content-Agnostic Moderation for Stance-Neutral Recommendations

Nan Li, Bo Kang, Tijl De Bie

TL;DR

This paper tackles the problem of recommendation systems amplifying user polarization by proposing "content-agnostic moderation." Instead of analyzing content, these methods adjust recommendations based solely on user-item interaction patterns, aiming to promote stance neutrality without risking censorship. Their novel methods, Random Dispersal and Similarity-Based Dispersal, show promise in balancing user exposure while maintaining recommendation quality.

What Problem Does This Paper Address?

Modern recommendation systems, prevalent in news and social media, often create "filter bubbles" and "echo chambers." By recommending content similar to what users have previously engaged with, these systems can inadvertently reinforce existing beliefs, leading to increased opinion polarization and societal fragmentation. While content moderation can help, it risks infringing on freedom of speech if it selectively suppresses content. This paper explores an alternative: content-agnostic moderation, which aims to reduce polarization by manipulating recommendation patterns rather than analyzing the content itself, thus avoiding direct censorship concerns.

Methods

The core of this paper is the introduction and evaluation of content-agnostic moderation techniques. These methods operate without access to the actual content of the items being recommended. Instead, they rely solely on relational properties, such as user-item interaction data (e.g., clicks, exposures).

The authors propose two novel methods:

  • Random Dispersal (RD) and Similarity-Based Dispersal (SD): These methods leverage biclustering of users and items based on aggregated recommendation data. They identify "tight clusters" of users and items and then swap less preferred items within a user's recommendations with items from other clusters. SD uses spectral embeddings for more targeted swaps, while RD uses random sampling.
  • Egalitarian Exposure Enforcement: This is a broader category of methods that aim for uniform exposure across different stances. The paper theoretically shows that if items are uniformly distributed across stances, then uniform exposure of these items will lead to stance neutrality.

The paper also theoretically demonstrates that content-agnostic moderation isn't guaranteed to work in all generic settings but can be effective under plausible assumptions, particularly when using relational proxies for stance information or when the data is favorably balanced.

Datasets

The paper does not use real-world datasets for training or evaluation. Instead, it introduces a novel simulation environment that includes a semi-synthetic dataset. This environment allows for controlled experiments with various user models, recommendation systems, and moderation strategies. The dataset is characterized by:

  • Stances: Defined as Left (L), Center (C), and Right (R).
  • User Preferences: A real-valued matrix reflecting the U.S. political landscape (56% L, 11% C, 33% R).
  • Item Stances: A binary matrix indicating the stance of each item.

The simulation includes four distinct scenarios to test robustness:

  1. Balanced Baseline: Equal distribution of items across stances.
  2. Bi-Polarized Content: High proportion of Left and Right items, low for Center.
  3. Propaganda Dominance: Heavily biased towards the Left.
  4. Amplified Right-Stance: User preferences are artificially skewed towards the Right.

Key Insights and Ideas

  • Content-Agnostic Moderation is Feasible (with caveats): While theoretically not guaranteed, the paper shows that by leveraging relational data, it's possible to significantly improve stance neutrality without direct content analysis. This is a crucial step towards less intrusive moderation.
  • Relational Proxies are Key: The effectiveness of content-agnostic moderation hinges on the recommender system's ability to capture user preferences and item stances through user-item interactions. If these interactions indirectly reflect stances, moderation can work.
  • Cluster Dispersal Outperforms Baselines: The proposed RD and SD methods are more robust and achieve better trade-offs between recommendation accuracy (CTR) and stance neutrality (JSD) compared to simpler methods like Round-Robin (RR) and Knapsack Constrained Optimization (KC), especially in challenging scenarios.
  • User Behavior Matters: User preference updates and the underlying media landscape (scenarios) significantly influence the effectiveness of moderation. Dynamic user models are essential for understanding long-term effects.
  • Trade-offs Exist, But Can Be Managed: Moderation strategies inherently involve a trade-off between recommendation quality (CTR) and stance neutrality. The paper demonstrates that methods like RD and SD allow for more controlled management of this trade-off, visualized through Pareto frontiers.

Key Results

  • Stance Neutrality Improvement: The proposed RD and SD moderation methods significantly improved stance neutrality (lower JSD values) across various scenarios, often outperforming baseline moderation techniques.
  • CTR Maintenance: RD and SD achieved these neutrality improvements with smaller drops in Click-Through Rate (CTR) compared to RR and KC, indicating better preservation of user engagement. For instance, in Table 3, RD and SD show CTR drops of around 12-13% while achieving better JSDs than RR and KC.
  • Robustness: The methods showed good performance across different scenarios, including balanced, bi-polarized, and propaganda-dominated content distributions.
  • Computational Efficiency: RD and SD were significantly faster than KC, operating at least 100 times faster, making them more practical for real-world deployment.
  • User Opinion Dynamics: Moderation, particularly RD and SD, helped temper user opinion extremeness (lower UMOE) and reduce polarization, especially in scenarios with extreme content.

Limitations and Future Work

  • Simulation-Based: The findings are based on a simulation environment and semi-synthetic data. While designed for realism, it may not fully capture the complexities of real-world platforms.
  • Single Topic Assumption: The current simulation assumes all items and interactions pertain to a single topic, simplifying the analysis but limiting its applicability to multi-topic recommendation systems.
  • User Model Simplification: While exploring static and dynamic user models, the authors acknowledge that more complex opinion updating models could be explored.
  • Hyperparameter Tuning: The paper notes that finer control over moderation strength through hyperparameter tuning is left for future investigation.
  • Real-World Deployment: Further research is needed to explore the impact of user mental state modeling and to validate these findings on real-world recommendation systems.

Summary updated

Abstract

Personalized recommendation systems often drive users towards more extreme content, exacerbating opinion polarization. While (content-aware) moderation has been proposed to mitigate these effects, such approaches risk curtailing the freedom of speech and of information. To address this concern, we propose and explore the feasibility of \emph{content-agnostic} moderation as an alternative approach for reducing polarization. Content-agnostic moderation does not rely on the actual content being moderated, arguably making it less prone to forms of censorship. We establish theoretically that content-agnostic moderation cannot be guaranteed to work in a fully generic setting. However, we show that it can often be effectively achieved in practice with plausible assumptions. We introduce two novel content-agnostic moderation methods that modify the recommendations from the content recommender to disperse user-item co-clusters without relying on content features. To evaluate the potential of content-agnostic moderation in controlled experiments, we built a simulation environment to analyze the closed-loop behavior of a system with a given set of users, recommendation system, and moderation approach. Through comprehensive experiments in this environment, we show that our proposed moderation methods significantly enhance stance neutrality and maintain high recommendation quality across various data scenarios. Our results indicate that achieving stance neutrality without direct content information is not only feasible but can also help in developing more balanced and informative recommendation systems without substantially degrading user engagement.
BibTeX
@inproceedings{li2025content,
  author = {Nan Li and Bo Kang and Tijl De Bie},
  title = {{Content-Agnostic Moderation for Stance-Neutral Recommendations}},
  booktitle = {ACM International Conference on Information and Knowledge Management},
  year = {2025},
  doi = {10.1145/3746252.3761111},
}
Figure