acceptodds
Under review as a conference paper at ICLR 2027

Socially-Aware Recommender Systems Mitigate User-Creator Features Clusterization

Abstract

Recommender systems shape online interactions by matching users with creators’ content to maximize engagement. Creators, in turn, adapt their content to align with users’ preferences and enhance their popularity. At the same time, users’ preferences evolve under the influence of both suggested content from the recommender system and content shared within their social circles. This feedback loop generates a complex interplay between users, creators, and recommender algorithms, which is the key cause of filter bubbles and polarization. We develop a social network-aware recommender system that explicitly accounts for this user-creators feedback interaction and strategically exploits the topology of the user's own social network to promote diversification. Our approach highlights how accounting for and exploiting users' social network in the recommender system design is crucial to mediate filter bubble effects while balancing content diversity with personalization. We show that a recommender system that greedly optimizes for user satisfaction leads to users' feature clusters formation. Ultimately, the proposed approach shows the power of socially-aware recommender systems in combating polarization and clusterization phenomena.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
recommender systems, dynamical systems, social networks

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Fgb2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper models closed-loop co-evolution of users, strategic content creators, and a recommender system via multi-dimensional Friedkin–Johnsen opinion dynamics over a social graph, proposing an unweighted 𝑑-hop neighborhood average to reduce feature clusterization.

Critical strengths

Clear Conceptual Formulation of Three-Way User–Creator–Recommender Co-Evolution Over a Social Graph

Critical weaknesses

Mathematical Flaws and Vacuousness in Theorem 2 (Appendix F.2) Lack of Baselines / Quantitative Real-Data Evaluation: No Comparison to Non-Social Diversification Baselines Incomplete ego-Facebook Evaluation & Scale-Invariant Silhouette Metric: On ego-Facebook, the paper provides only a qualitative 3D scatter plot at 𝑡=20 (Figure 4) with no quantitative satisfaction or clusterization curves. Moreover, measuring global clusterization solely via the 𝑘 -means silhouette coefficient (Eq. 8–9, where 𝑘 is re-selected to maximize silhouette) is scale-invariant and can report high clusterization even when all users have converged into an arbitrarily small ball around consensus.

Questions for authors

While studying the three-way feedback loop between socially connected users, adaptive creators, and a recommender system is well-motivated, the paper's core theoretical results (Theorem 2) contain critical mathematical errors (a vacuous threshold 𝐹>8𝑛 on [−1,1]^𝑛, a non-positive denominator when 𝜆_𝑖=0 and conflation of ∥𝑥∥_2 with ∥𝑥∥22∥𝑥∥_2^2 , the proposed 𝑑-hop unweighted neighborhood average collapses to a global mean in small-world graphs by d=3, and the empirical evaluation lacks baselines and quantitative metrics on the ego-Facebook graph. Can you explain?

Reviewer_EbYsOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies a feedback loop between a recommender system, users connected by a social network, and adaptive content creators. A possible solution is proposed that suggests to users items similar to the average preferences of users in their neighborhood, potentially beyond direct social ties. The neighborhood radius can be used to tune a tradeoff between personalization and reduced clustering. The theory examines conditions under which clustering emerges, while simulations over synthetic networks and a Facebook social network suggest that the proposed approach can potentially reduce clustering with still relatively high satisfaction.

Critical strengths
  • Relevant problem, taking into account influence among users as well as adaptation of content creators.
  • Simple recommendation rule, with an understandable parameter for adjusting the neighborhood considered.
  • Interesting simulations that highlight the satisfaction–clustering tradeoff, including how the results change with social network density and the size of the recommendation shortlist.
Critical weaknesses
  • Theorem 2 seems to me applicable only assuming no influence among different users, since (A) is diagonal. This makes it less relevant to the paper’s main setting. Moreover, the separation condition involving Equation 20 appears impossible to satisfy within the bounded feature space. If I understood correctly, assuming a positive denominator, the equation implies (F\geq24n), whereas the difference of squared distances is at most (4n). This seems to make the condition vacuous as written.
  • Corollary 1 is an immediate consequence of comparing distances at a given state. As such, it does not tell much about the original setting, in which creators and users evolve over time. It is not clear to me whether greedily recommending what maximizes current satisfaction is indeed optimal wrt the given satisfaction objective, which averages over time.
  • The evaluation is entirely simulation based. Although the Facebook network is real, user preferences, creator features, and their behavioral dynamics are simulated. This limits what the results establish about a real-world deployment.
  • Most comparisons seem to be obtained by changing the neighborhood radius in the proposed approach. This does not establish how much the improvement depends on meaningful social connections rather than averaging more generally.
  • The variance reported in the graphs is computed across users. This does not show how stable the findings are across independent simulation runs.
Questions for authors
  1. How can we distinguish between users becoming more similar to each other in their preferences or average consumption, and each user seeing a more diverse set of items? Lower inter-user diversity, i.e. more homogenization, is not the same thing as higher intra-user diversity, i.e. broader individual consumption. Here’s a recent paper on this topic that might be worth looking at, Filter Bubble or Homogenization? (WWW ’24) by Anwar et al.
  2. Could you include more baselines to show the importance of taking the actual social network into account, compared with, e.g., a shuffled network, a random neighborhood, or the average preferences of the overall population? Comparing these at similar satisfaction levels would be useful.
  3. Can you provide a nontrivial parameter setting satisfying Theorem 2’s separation condition, or clarify whether Equation 20 needs correction? Also, what additional argument supports Corollary 1 for the cumulative satisfaction objective?
Reviewer_dBkZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper explores recommendation phenomena on social networks, arguing that in user-creator scenarios, past work overlook the user-user social influence. A framework is built on top of this existing work to update the internal preferences of both entity types from a more complete perspective and mitigate filter bubbles arising from recommender systems that focus solely on user utility.

Critical strengths
  • The paper tackles a topic of interest for certain industrial scenarios: social network modeling with multiple entity types.
  • The paper jointly considers user-creator influence and user-user social influence, addressing filter-bubble mitigation under evolving user preferences and creator content.
  • The paper shows how its approach can find a trade-off between user satisfaction and polarization, as measured by clustering in the simulated feature space. It provides useful ablations on the impact of parameters k (Appendix) and d-hops over multiple time cutoffs.
Critical weaknesses
  • The novelty appears quite limited. The framework resembles a heterogeneous network model with two entity types instantiated in a user–creator scenario with user-user, user-creator, and creator-user edge types. Its weighted aggregation and feature updates are conceptually related to existing graph-based representation-learning approaches, although they model temporal dynamics rather than gradient-based learning. If we consider creators as items, related approaches can be found in the literature, such as DiffNet++. The paper should clarify its methodological novelty relative to these approaches.
  • The approach to reducing user clusterization appears to leverage a smoothing effect analogous to oversmoothing in graph neural networks, where entity-specific information can be lost. This makes it difficult to view the finding as new or unexpected beyond the user-creator formulation. Additionally, the neighborhood aggregation uses no distance-based weight decay, meaning that nearby and distant users within the selected neighborhood contribute equally to the recommendation reference. Further explanation of this design choice, or a comparison against a distance-decay-weighted variant, would be useful.
  • A similar concern applies to the user satisfaction formulation. A user’s satisfaction with the recommended creators depends on all previous time steps, with every time step contributing equally and no forgetting effect. A recency-weighted formulation, potentially varying by user, would be worth discussing (different users may have different forgetting rates) and would clarify whether the trade-off persists at later stages.
  • Both the synthetic and real-dataset environments involve simulated dynamics in really small graphs. Even for the real Facebook dataset creator entities are introduced with randomly generated initial features. Additionally, complex user and creator representations seem to be represented with too low dimensional vectors. The paper would benefit from validation on a larger real user–creator environment, ideally using observed longitudinal interactions and a larger feature set.
  • In practice, content creators are often also users, meaning that they can be influenced by other content creators. They may also create varied content targeting different audiences. A temporally evolving heterogeneous graph with users, content items, and potentially brands/creators could enrich the framework with additional feedback signals.
  • The paper still lacks several useful ablations and comparisons. The current evaluation primarily compares variants of a single recommendation rule, using its greedy special case as the baseline and neighborhood smoothing to reduce polarization. Ablations involving different entity behaviors (e.g. frozen creators or frozen user preferences) along with alternative diversity/exploration approaches, neighborhood perturbations, and other recommendation policies would substantially strengthen the paper’s contributions and positioning within the recommendation literature.

Note: The paper has redundant explanations that could be compacted to include insightful experiments. Additionally, it contains multiple typos and errors in the text and equations worth fixing e.g. Equation 11b is missing subindex j that appears in Equation 5, "sinthetic", "n-dmensional", "feautres", "preferemces", among others.

Questions for authors

How does this approach behave against alternative diversity-aware methods, recommender system backbones and forgetting effects?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.