Simulations of collaborative-filtering recommender systems show that iterated retraining on their own recommendations progressively contracts the diversity of content shown to each user, trapping individuals in echo chambers, and three attempted optimization approaches all failed to let a single user widen their own recommendations by manipulating only their rating vector. The unresolved core is whether an efficient algorithmic procedure for such single-user escape exists, or can be proven not to exist; the further questions of whether diversity contraction survives realistic user-behavior models and whether exploration-based mitigations work in dataset-grounded settings remain unreplicated, but they are not part of that core.

Background

Collaborative filtering (CF) is the dominant technique behind commercial recommenders: it predicts a user's ratings for unseen items from the ratings of similar users, most commonly via matrix factorization — the singular-value decomposition (SVD) family of models, which represent users and items as low-rank latent vectors whose inner products give predicted ratings. Because a deployed recommender is periodically retrained on the interactions it itself provoked, it operates in a feedback loop: today's recommendations shape tomorrow's training data. A recurring concern, theorized by Jiang et al. (AIES 2019) under the name degenerate feedback loops and simulated by Chaney et al. (RecSys 2018) as algorithmic confounding, is that this loop progressively narrows the range of content each user sees — an echo chamber or filter bubble effect. Noordeh, Levin, Jiang, and Shadmany (arXiv:2011.03890, 2020) gave this concern a quantitative, dataset-grounded form. They trained an SVD recommender (the surprise library implementation) on MovieLens — the standard public movie-rating benchmark — and simulated epochs in which each user receives the top-kk predicted movies, and, crucially, accepts every recommendation and re-enters exactly the rating the model predicted. Content diversity was measured in MovieLens tag genome space: each movie is a vector of community-derived tag relevances (Vig, Sen, and Riedl, ACM TiiS 2012), and a user's diversity is the spread of their recommended movies in that space. Across configurations (10–400 epochs, 12–100 recommendations per epoch, 10–1000 users), recommended-content diversity contracted sharply — a 7σ7\sigma drop within 10 iterations in one simulation — and the qualitative effect was insensitive to hyperparameters. The authors then asked the natural individual-agency question: can a lone user, confined to editing their own rating vector, break out? They framed widening one's recommended diversity as an optimization problem whose objective is a black-box function of one's ratings (it requires retraining and re-ranking the whole model), and tried finite-difference/trust-region methods (local optimization that estimates the objective's gradient from small perturbations of the ratings), derivative-free sampling optimizers, and two heuristics (rate maximally distant movies; rate random movies highly). None significantly increased the objective. The paper's conclusion states escape is 'infeasible', but this rests solely on those failed empirical attempts — no impossibility or complexity result is given. Related literatures stop short of the question: shilling and data-poisoning attacks manipulate recommenders to promote particular items rather than the attacker's own diversity; the theoretical analysis of degenerate loops by Jiang et al. holds only in a deliberately simplified model and addresses system-side exploration, not single-user action; and subsequent simulation frameworks with more realistic user behavior (partial acceptance, stochastic or satisfaction-dependent ratings) have not re-run the SVD/MovieLens/tag-genome contraction experiment.

Problem Statement

The problem is to settle the question about diversity contraction in iterated collaborative-filtering feedback loops that the echo-chamber simulation study of Noordeh et al. (2020) leaves open: whether a single user, confined to editing their own rating vector, can escape the contraction. The source poses two central questions — whether recommender systems substantially reduce content diversity and contribute to echo-chamber formation over time, and, if so, whether it is possible for a user to break out of the echo chamber by manipulating solely their own content ratings — answers the first affirmatively by simulation, and addresses the second only by reporting that three black-box optimization approaches failed while asserting infeasibility in its conclusion.

Decide algorithmically whether a single user can escape: does there exist an efficient procedure by which a user, manipulating only their own rating vector, reliably and substantially increases the diversity of the content recommended to them by an SVD-type collaborative-filtering system retrained on the augmented data — or is there a formal sense (computational complexity, information-theoretic, or learning-theoretic) in which no such procedure exists? The source reports only empirical failure of three optimization approaches; the answer is either such a constructive procedure or a proof that no such procedure exists.

Both branches concern the source's retrain-and-recommend pipeline, in which the acting user's strategic ratings are appended to the training data and the model is refit and re-ranked, so that the manipulation changes the very recommended set whose diversity is measured; an impossibility or hardness result is therefore to be stated for a formalization that retains this coupling (a variant in which the acting user's appended ratings have no influence on the recommended set — for instance a fixed-historical-data variant or an infinite-user limit — does not address the question). Within that setting, "efficient" is understood in the standard sense of running in time polynomial in the size of the instance (users, items, observed ratings), counting each retraining-and-re-ranking evaluation of the black-box diversity objective as one oracle call; "substantially" means an increase in that user's recommended-content diversity DD — the mean pairwise Euclidean distance between the recommended movies in tag-genome space, as in the source — relative to the otherwise identical simulation without the strategic manipulation, beyond run-to-run variation across seeds; and "reliably" means that the increase is not an artifact of a particular seed, initialization, or perturbation draw. Because the source poses the question for "a user", an answer also states explicitly the set of users for which escape is established — one user, a specified subset, or all users of the simulated population.

Related questions that a resolution of the above need not settle, retained here because they condition how any answer is to be interpreted: whether the diversity-contraction effect survives realistic user-behavior models — partial acceptance of recommendations, stochastic rating behavior, or feedback depending on user satisfaction rather than the predicted score — in the dataset-grounded SVD/MovieLens tag-genome framework, replacing the source's exact-acceptance assumption; whether the system-side mitigations suggested by prior theory (explicit random exploration, enlarged candidate pools) preserve diversity in such realistic dataset-grounded settings, and with what trade-off against recommendation accuracy; and how far the contraction effect extends beyond the specific SVD implementation and MovieLens tag-genome setup, across recommender architectures and interaction regimes.

References

  1. Emil Noordeh, Roman Levin, Ruochen Jiang, Harris Shadmany, 'Echo Chambers in Collaborative Filtering Based Recommendation Systems', 2020. arXiv:2011.03890, DOI 10.48550/arXiv.2011.03890, https://arxiv.org/abs/2011.03890
  2. Ray Jiang, Silvia Chiappa, Tor Lattimore, András György, Pushmeet Kohli, 'Degenerate Feedback Loops in Recommender Systems', 2019. arXiv:1902.10730, DOI 10.48550/arXiv.1902.10730, https://arxiv.org/abs/1902.10730
  3. Allison J. B. Chaney, Brandon M. Stewart, Barbara E. Engelhardt, 'How Algorithmic Confounding in Recommendation Systems Increases Homogeneity and Decreases Utility', 2018. arXiv:1710.11214, DOI 10.48550/arXiv.1710.11214, https://arxiv.org/abs/1710.11214
  4. Yingqiang Ge, Shuya Zhao, Honglu Zhou, Changhua Pei, Fei Sun, Wenwu Ou, Yongfeng Zhang, 'Understanding Echo Chambers in E-commerce Recommender Systems', 2020. arXiv:2007.02474, DOI 10.48550/arXiv.2007.02474, https://arxiv.org/abs/2007.02474
  5. Mihaela Curmei, Andreas Haupt, Benjamin Recht, Dylan Hadfield-Menell, 'Towards Psychologically-Grounded Dynamic Preference Models', RecSys 2022. arXiv:2208.01534, DOI 10.48550/arXiv.2208.01534, https://arxiv.org/abs/2208.01534 (also DOI 10.1145/3523227.3546778)
  6. Erica Coppolillo, Simone Mungari, Ettore Ritacco, Francesco Fabbri, Marco Minici, Francesco Bonchi, Giuseppe Manco, 'Algorithmic Drift: A Simulation Framework to Study the Effects of Recommender Systems on User Preferences', 2024. arXiv:2409.16478, DOI 10.48550/arXiv.2409.16478, https://arxiv.org/abs/2409.16478
  7. Tim Faverjon, Jean-Philippe Cointet, Pedro Ramaciotti, 'How do recommender systems learn political opinions? A semi-synthetic step-by-step experiment', PLOS ONE, 2026. DOI 10.1371/journal.pone.0349341, https://doi.org/10.1371/journal.pone.0349341
  8. Edoardo Bianchi, 'Exploration on Demand: From Algorithmic Control to User Empowerment', 2025. arXiv:2507.21884, DOI 10.48550/arXiv.2507.21884, https://arxiv.org/abs/2507.21884
  9. Difu Feng, Qianqian Xu, Zitai Wang, Cong Hua, Zhiyong Yang, Qingming Huang, 'Quantifying the Potential to Escape Filter Bubbles: A Behavior-Aware Measure via Contrastive Simulation', 2025. arXiv:2512.03067, DOI 10.48550/arXiv.2512.03067, https://arxiv.org/abs/2512.03067
  10. Ryoma Sato, 'Towards Principled User-side Recommender Systems', 2022. arXiv:2208.09864, DOI 10.48550/arXiv.2208.09864, https://arxiv.org/abs/2208.09864
  11. Sarah H. Cen, Andrew Ilyas, Aleksander Madry, 'User Strategization and Trustworthy Algorithms', 2023. arXiv:2312.17666, DOI 10.48550/arXiv.2312.17666, https://arxiv.org/abs/2312.17666
  12. Jiwen Zhou, Xiang Liu, Mingming Li, Pengbo Mo, Jiao Dai, Honglei Lv, Jizhong Han, Songlin Hu, 'Can We Steer the Black-Box? Towards Controllability-Centric Evaluation of Recommender Systems with Collaborative Agents', 2026. arXiv:2607.13418, DOI 10.48550/arXiv.2607.13418, https://arxiv.org/abs/2607.13418
  13. Luca Pappalardo, Salvatore Citraro, Giuliano Cornacchia, Mirco Nanni, Valentina Pansanella, Giulio Rossetti, Gizem Gezici, Fosca Giannotti, Margherita Lalli, Giovanni Mauro, Gabriele Barlacchi, Daniele Gambetta, Virginia Morini, Dino Pedreschi, Emanuele Ferragina, 'A survey on the impacts of recommender systems on users, items, and human-AI ecosystems', 2024. arXiv:2407.01630, DOI 10.48550/arXiv.2407.01630, https://arxiv.org/abs/2407.01630
  14. Jesse Vig, Shilad Sen, John Riedl, 'The Tag Genome', ACM Transactions on Interactive Intelligent Systems 2(3), 2012. DOI 10.1145/2362394.2362395, https://doi.org/10.1145/2362394.2362395