Content-Agnostic Moderation for Stance-Neutral Recommendation

TL;DR

The paper addresses polarization risk in personalized recommendations and fairness concerns around content moderation by proposing content-agnostic moderation that relies solely on relational properties. It develops a theoretical non-determinacy result showing neutrality cannot be guaranteed in full generality, yet identifies practical conditions under which content-agnostic moderation can work via proxy-based moderation and egalitarian exposure enforcement. It introduces two novel cluster-dispersal moderation methods, RD and SD, and validates them in a reusable simulation framework across diverse data scenarios, demonstrating improved stance neutrality with manageable CTR loss and faster computation than baselines. The work offers a proof-of-concept pathway to reduce polarization without content-level censorship, highlighting practical implications for deploying moderation in real-world recommender systems and outlining avenues for richer modeling and empirical validation.

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.

Figures (5)

• Figure 1: The feedback loop of the simulation framework.
• Figure 2: Pareto frontiers of different moderation strategies with Oracle-CB recommender and user preference updating in S2 and S3. Hyperparameters are bracketed, the meanings of which are explained in paragraph \ref{['sec:hyperparameters']}. Markers are centered at the mean values over multiple runs, with errorbars roughly proportional to the standard deviations.
• Figure 3: User opinion and neighborhood stance distribution at $t_0$ and $t_{60}$ without/with moderation of varied strengths.
• Figure 4: Pareto frontiers of different moderation strategies with Oracle-CB recommender and user preference updating. Hyperparameters are bracketed, the meanings of which are explained in paragraph \ref{['sec:hyperparameters']}. Markers are centered at the mean values over multiple runs, with errorbars roughly proportional to the standard deviations.
• Figure 5: Pareto frontiers of different moderation strategies with Oracle-CB recommender and without user preference updating.

Theorems & Definitions (3)

• definition 1: Problem setup
• definition 2: Content-Agnostic Moderation Function
• theorem 1: Non-Determinacy of Content-Agnostic Moderation