acceptodds
Withdrawn from ICLR 2027

Feature Diversity Is Not Monotone in Content Exploration: Evidence from Co-Adapting User–Creator Recommender Systems

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recommender systems influence both users, whose preferences can change with the content they see, and creators, who may adapt their content to the audiences they reach. We study how exploration affects recommendation diversity, the variety of content recommended to each user, while users remain influenced by their underlying interests and creators by their established specializations. We establish sufficient conditions under which intermediate exploration yields higher long-run recommendation diversity than both uniform recommendation, which splits each user's exposure equally across creators, and rates close to pure exploitation, which concentrates exposure on each user's highest-scoring creator. We also establish conditions for convergence to a unique equilibrium when exploration is sufficiently close to uniform. Simulations with synthetic vectors and representations learned from MovieLens show that intermediate exploration can improve recommendation diversity, with the benefit varying with the initialization of the dynamics and how strongly users and creators respond to recommendations.

Primary area:
learning on time series and dynamical systems
Keywords:
recommender systems, recommendation diversity, co-adaptive dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_bRPtOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper conducts simulations to study how exploration affects recommendation diversity, measured by the variance of content features recommended to each user, while users remain influenced by their underlying interests and creators by their established specializations. However, the key motivation, contributions, and findings of this paper are not clear and well illustrated. Overall, I think this paper is in very poor presentation and organization, and is hard to understand its main points.

Critical strengths

To be honest, the only strength I can tell from the paper is extensive theoretical analysis I guess.

Critical weaknesses
  1. The main motivation, contribution and the most significant findings from the experiments are not clear and well present?

  2. Experiments are only conducted in one real-world dataset, which is not enough to draw any conclusions.

  3. The related work is in very poor presentation, from which I cannot judge the positioning of this paper.

Questions for authors
  1. Why use variance of content features to evaluate recommendation diversity, instead of more intuitive metrics such as catagory coverge?

  2. Figure 1 is over-simple, from which I cannot figure out the pipeline of this paper. Could you please provide a more detailed and well-structed diagram?

  3. What's your positioning against other diversity-based recommendation methods?

Reviewer_pMEnOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The model represents MM users and NN creators in the same dd-dimensional feature space. At each step, users move toward the exposure-weighted average of the creators recommended to them, and creators move toward the average preferences of the users they reach. Fixed user and creator anchors limit both responses. Exposure follows an ϵ\epsilon-greedy rule: a fraction 1−ϵ\epsilon is assigned to each user’s highest-scoring creator, and the remaining fraction is distributed uniformly across creators. Recommendation diversity (RD) is the average, across users, of the exposure-weighted variance of creator content vectors. The experiments test 13 values of ϵ\epsilon using synthetic vectors and vectors fitted to MovieLens 1M. An interior grid value gives the largest long-run RD in all four synthetic settings and three of the four MovieLens settings. Uniform exposure gives the largest value in the MovieLens setting with the strongest user response and weakest creator response.

Critical strengths

1. The three reported outcomes are not conflated: The paper reports recommendation relevance (RR), creator diversity (CD), and recommendation diversity (RD) separately. Positive CD does not imply positive RD. Under pure exploitation, each user receives content from one creator and therefore has RD=0, even if the creator vectors are widely separated. This separation is necessary for the stated research question, which concerns variation within a user’s recommendation distribution rather than variation among all creators on the platform.

2. The frozen-state calculation identifies one source of the interior maximum: The formula RD(z,ϵ)=ϵVall(z)+ϵ(1−ϵ)Vtop(z)RD(z,\epsilon) = \epsilon V_{\mathrm{all}}(z) + \epsilon(1-\epsilon)V_{\mathrm{top}}(z) gives the exact frozen-state condition Vtop>VallV_{\mathrm{top}}>V_{\mathrm{all}} for a strict interior maximum. It also separates two effects of the exposure rule: variance under uniform creator sampling and the additional between-component variance obtained by mixing the uniform distribution with each user’s highest-scoring creator. The dynamic and frozen-state comparisons then quantify how much state adaptation changes this static effect.

Critical weaknesses

1. The experiments do not show that co-adaptation is needed for non-monotonicity:

Prior work already treats time-varying user factors, recommendation-induced preference change, creator responses to exposure, and mutual user–creator influence in a shared feature space. Relative to that literature, this paper adds fixed anchors and analyzes the resulting model under ϵ\epsilon-greedy exposure. The main qualitative result is also present without state adaptation. In each of the seven settings whose dynamic curve has an interior grid maximum, the corresponding frozen-state curve has an interior maximizer. The simulations therefore show that adaptation changes the size and location of the gain; they do not show that adaptation creates the interior maximum. A stronger result would identify parameter settings in which the frozen-state curve is monotone but the dynamic curve is not, or give conditions under which adaptation changes the sign of the derivative.

2. The convergence theorem does not cover the exploration rates at which the largest experimental gains occur:

Theorem 4.1 applies only when ϵ\epsilon lies in a narrow interval below one. For the experimental configurations, the reported interval widths are approximately 3.7×10−113.7\times10^{-11} to 7.2×10−77.2\times10^{-7}. All nonuniform values in the 13-point experimental grid lie outside these intervals. Several reported grid maxima occur at epsilon=0.35epsilon=0.35 or ϵ=0.65\epsilon=0.65. At those values, the numerical checks support convergence from the simulated initial states, but the theorem does not establish a unique equilibrium or convergence from every admissible initialization. The local result near ϵ=1\epsilon=1 therefore does not account for the parameter region in which the largest gains are reported.

Questions for authors
  1. Is there a parameter setting in which the frozen-state function RD(zunif∗,ϵ)RD(z_{\mathrm{unif}}^*,\epsilon) is monotone on [0,1], but the long-run dynamic RD has a strict interior maximum? If so, please include it. This would show that adaptation can create, rather than only alter, the non-monotonicity.

  2. Corollary B.3 determines the sign of the creator-adaptation term when response fractions are common within the user population and within the creator population. Can the sign be characterized for heterogeneous response fractions in equations (29)–(31)? If not in full generality, are there sufficient conditions stated in terms of the anchors or response coefficients?

  3. What makes the interval in Theorem 4.1 so small in the reported experiments? Please separate conservatism introduced by the proof bounds from actual changes in the number or stability of equilibria. Equilibrium continuation in ϵ\epsilon, local Jacobian spectral radii, or intervals over which a fixed assignment remains self-consistent would help answer this.

  4. Outside the interval covered by Theorem 4.1, do widely separated initial states converge to the same fixed point or periodic orbit? Please report this for the settings and values of ϵ\epsilon that produce the largest RD. If different initial states reach different attractors, the distribution of resulting RD values is more informative than a single trajectory.

  5. Does the interior maximum persist for a diversity measure that is not forced to zero under pure exploitation? Relevant checks include exposure entropy, catalog coverage, or the average pairwise distance among creators shown to a user over a time window.

Reviewer_nLq3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes PsyCAT for self-report computerized adaptive testing. It combines explicit response-style modelling, Gaussian posterior mixtures, shared-attention post-calibration, minimax aggregation, and information-regularized item selection. Evaluation includes held-out response prediction on six questionnaire banks, controlled latent-change experiments, and an expert study of information leakage.

Critical strengths
  • The paper addresses response-style variation and information leakage in self-report CAT through an integrated estimation and selection framework.
  • PsyCAT achieves the lowest mean endpoint RPS among the evaluated methods across six questionnaire banks, providing consistent evidence of improved response prediction.
Critical weaknesses
  • The paper addresses response-style variation and information leakage in self-report CAT through an integrated estimation and selection framework.
  • PsyCAT achieves the lowest mean endpoint RPS among the evaluated methods across six questionnaire banks, providing consistent evidence of improved response prediction.
Questions for authors

Please see the weakness.

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

Related papers

Discussion (0)

Sign in to comment.