Pith. sign in

REVIEW 3 major objections 4 minor 60 references

Two new datasets pair user-preference data with community content warnings, and a first analysis with them shows personalized recommenders deamplify, not amplify, sensitive content.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Two benchmark datasets link user-preference data (MovieLens, Archive of Our Own) with community content-warning labels to study sensitive-content exposure in recommender systems.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection The two datasets are a real asset, but the headline amplification result is currently undercut by the metric's zero-history handling. the 3 major comments →

arxiv 2509.07269 v1 pith:MZMA6SYA submitted 2025-09-08 cs.IR cs.AI

Datasets for Navigating Sensitive Topics in Recommendation Systems

Amelia Kovacs , Jerry Chee , Kimia Kazemian , Sarah Dean This is my paper
classification cs.IR cs.AI
keywords recommendation systemscontent warningsalgorithmic amplificationsensitive contentdatasetsMovieLensArchive of Our Owntrigger warnings
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Recommendation systems are typically judged by engagement, but a growing public concern is that they expose people to sensitive or harmful content they would rather avoid. The paper's central move is to make that concern measurable: it builds two datasets that attach community-generated content warnings to standard preference data — MovieLens ratings matched to Does the Dog Die? vote totals, and Archive of Our Own kudos history matched to a 36-category trigger-warning taxonomy. On top of the datasets it defines a Warning Amplification metric (the share of warned items in a user's top-k recommendations divided by the share in that user's history), then runs a first pass with three recommenders. The headline result is that the personalized algorithms deamplify warnings relative to the Random and TopPop baselines, even though several warning categories correlate with popularity. If it holds, the research community gains reusable benchmarks for recommender harm, plus a first signal that personalization trained on engagement alone does not push warning-dense content at users in these two domains.

Core claim

The interplay between personalized recommendation and sensitive content can be measured with warning-labeled interaction data, and the paper supplies the data: ML-DDD pairs MovieLens ratings with 137 Does the Dog Die? warnings, and AO3 pairs fan-fiction interactions with 36 trigger-warning categories. With a new Warning Amplification metric in hand, it reports a preliminary result: across 1,000 users, collaborative filtering (SVD and ALS) amplifies warnings less than the Random and TopPop baselines — personalization deamplifies sensitive content relative to baselines. The paper offers both datasets as reusable benchmarks and the deamplification finding as an opening.

What carries the argument

The load-bearing objects are the two sensitivity-labeled datasets and the metric that connects them to algorithm behavior. Warning Amplification@k is defined, for a given user and warning, as the fraction of warned items in the user's top-k recommendations divided by the fraction of warned items in the user's interaction history, normalized by subtracting 1 (with a hallucinated single warned item in history when the denominator would be zero). The datasets rest on a binarization rule: a work is labeled 'Clear Yes' for a warning when at least 75% of community votes are Yes, 'Clear No' when at least 75% are No, and ambiguous otherwise; AO3 works are labeled by presence or absence of warnings i

Load-bearing premise

The load-bearing premise is the labeling rule: a work counts as carrying a warning when at least 75% of community votes say yes, with no minimum vote count, so a work with three yes-votes and one no-vote becomes a confident 'Clear Yes' label; for the AO3 half, an added premise is that the roughly 30% of works with collected interactions represent the corpus. All permutation tests and amplification scores inherit whatever error these choices create.

What would settle it

Re-run the entire amplification pipeline with the warning labels re-binarized at a 50% majority threshold (with and without a minimum-vote floor, e.g., at least 10 votes for ML-DDD). If under those labels any personalized algorithm amplifies warnings relative to its Random or TopPop baseline — or if the ordering of the algorithms flips — then the paper's deamplification conclusion is an artifact of the 75% rule, not a property of personalization. A companion check: hand-label a random sample of works the rule binarized as Clear Yes/Clear No and measure agreement, especially for low-vote items;

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Researchers can evaluate recommenders on a harm axis — unwanted exposure to warning-labeled content — rather than on engagement alone, using two public, reproducible benchmarks.
  • If the deamplification result is robust, engagement-maximizing personalized recommenders are not the main driver of warning-dense exposure in these domains; non-personalized popularity ranking (TopPop) is the larger amplifier in movies.
  • Multi-objective recommender work that co-optimizes engagement with a harm proxy gains concrete datasets to train and test on.
  • The documented warning–popularity correlations (blood/gore and kidnapping negatively correlated with movie ratings; pornography positively correlated with kudos) give designers a prior on which categories need active deamplification.
  • The datasets open the paper's own listed questions to empirical work: label disagreement in DDD votes, modeling negative preferences, and designing user controls for sensitive content.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The deamplification verdict may ride on popularity, not personalization: in ML-DDD, TopPop amplifies most precisely because warning-dense movies are popular, while the latent-factor model under-recommends them without any explicit knowledge of warnings. A popularity-matched random baseline would separate the two effects.
  • A natural stress test is to re-run the pipeline at a 50% majority-vote threshold with a minimum-vote floor; if the algorithm ordering flips, the 75% rule is carrying the conclusion rather than the recommenders.
  • Keeping vote counts instead of binarized labels would let the community model warning confidence and per-user avoidance preferences, turning label disagreement itself — which the paper notes as future work — into a signal for user agency.
  • The same construction could be ported to platforms with deployable sensitive-content controls (social feeds with content filters), turning the benchmark into an audit tool for whether real systems change exposure at all.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces two novel datasets for studying sensitive content in recommender systems: ML-DDD, which combines MovieLens 25M ratings with content warnings from Does the Dog Die?, and AO3, which combines Archive of Our Own kudos interactions with trigger warnings from the Webis Trigger Warning Corpus. Each dataset contains a sensitivity table of item-warning labels and an interaction table of user-item ratings/kudos. The authors provide descriptive analyses of the relationship between warnings and popularity/user preferences using permutation tests, and a preliminary amplification analysis comparing Random, TopPop, and personalized matrix-factorization recommenders (SVD for ML-DDD, ALS for AO3) via a novel Warning Amplification@k metric. The central claim, stated in §6.3, is that personalized algorithms do not amplify warnings relative to non-personalized baselines.

Significance. The datasets address a real gap: there are few resources that combine standard preference data with community-generated content-warning labels, and the artifacts are publicly released with transparent construction steps. The descriptive analyses are genuine null-hypothesis tests on external data sources, and the permutation-test results are a useful starting point. The amplification analysis, however, is not currently supported by the proposed metric due to a load-bearing design choice, so the headline empirical conclusion should not be accepted without revision. If the metric is repaired and the label-threshold robustness is established, the datasets themselves could be a valuable community benchmark.

major comments (3)
  1. [§5.2, Warning Amplification Metric] The hallucination rule is load-bearing and can manufacture the reported deamplification. For a user with zero warning items in history, the denominator becomes 1/(|H|+1), so Amp@k = r(|H|+1) − 1, where r is the warning fraction in recommendations. Whenever r < 1/(|H|+1)—which occurs for rare warnings and short histories—the metric reports negative amplification even though the user moved from 0% warning exposure in history to r>0 in recommendations. Since most users have zero history for most warnings, these artificial negative scores can dominate averaged amplification and directly produce the §6.3 conclusion. No sensitivity analysis over the hallucination count is given. Please re-analyze excluding zero-history users, or use a well-defined baseline such as a global item-warning prevalence, and report results under alternative rules.
  2. [§3.1, Clear Yes / Clear No labeling] The binarization rule—Clear Yes if ≥75% of votes are Yes, Clear No if ≥75% are No—has no floor on total votes. A work with 3 Yes and 1 No is labeled Clear Yes. This label is then used in every permutation test and amplification calculation, so any systematic error from low-vote items propagates throughout the analysis. The paper does not validate the binarized labels against any ground truth, nor does it report vote-count distributions or sensitivity to the 75% threshold or to a minimum-vote floor. Please add such robustness checks, e.g., rerun key analyses with a minimum vote count (e.g., ≥5 or ≥10) and with alternative thresholds.
  3. [§3.2, AO3 subset construction] The AO3 dataset collects interactions for nearly 30% of the Webis corpus, but the paper does not analyze how this subset was selected or whether it is representative. The filtering rules (exclude works with <3 public kudos and users with <3 interactions) likely bias the subset toward popular, heavily-kudosed works. This selection bias could shift all AO3 summary statistics, permutation tests, and amplification results. Please compare the collected subset to the full Webis corpus on available metadata (e.g., hit counts, kudos counts, tag frequencies) and report the selection mechanism or a representativeness analysis.
minor comments (4)
  1. [Throughout] Several typos: 'recieved' (§2.1), 'recommeneded' (§6.2), 'Low Left' should be 'Bottom Left' in Figure 1 caption. A proofreading pass is recommended.
  2. [§5.2] The notation for Amplification@k is introduced informally; please provide an explicit equation with symbols for history size, recommendation size, and the hallucination step.
  3. [§6.2.1, Table 2] Precision, recall, and F1 values are extremely small (e.g., 1e-4 to 1e-2). This is common in sparse recommendation settings, but a brief note contextualizing these magnitudes would help readers avoid misinterpreting them.
  4. [§3.2, Figure 3] The caption says 'More users gave higher average ratings to the sensitivity label: 57.30% for ML-DDD and 72.60% for AO3,' but the text says 57.30% fall below the red line. Please ensure the caption and text are consistent about the direction of the effect.

Circularity Check

0 steps flagged

No circular derivation: datasets and preliminary analyses are self-contained; the amplification metric caveat is a validity concern, not circularity.

full rationale

The paper's central contributions are two datasets assembled from external sources (MovieLens + Does the Dog Die?, Webis Trigger Warning Corpus + AO3) with explicit vote-threshold labeling rules. The descriptive analyses are permutation tests against null hypotheses, and the recommender evaluation trains SVD/ALS with hyperparameters tuned to RMSE/recall, not to the amplification outcome. The warning-amplification metric is defined independently of the algorithms' outputs; although the zero-history handling ('hallucinate a single item...') can bias the sign of Amp@k for users with no prior warning, this is an arbitrary imputation choice rather than a fitted parameter or definitional equivalence, and the comparison between personalized and baseline algorithms is an empirical result. The only self-citations ([7], [8]) appear in related-work context and do not carry the paper's argument. No load-bearing step reduces to its own input, so no significant circularity found.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The central artifact is a data linkage, so the ledger records the labeling and sampling choices that determine dataset validity: the 75% binarization threshold, the hallucinated-history rule in the amplification metric, evaluation horizons and subset sizes, and the domain assumptions that community warnings and identifier matches are trustworthy. No new physical or conceptual entities are postulated; the Warning Amplification metric is a measurement definition rather than an entity.

free parameters (5)
  • Clear Yes / Clear No vote threshold = 75%
    Hand-chosen threshold in Section 3.1 to binarize DDD vote shares into warning present or absent; no minimum vote count and no validation, so labeling errors propagate into every downstream statistic.
  • Hallucinated history item count = 1
    Section 5.2 inserts one synthetic warning item into any user history that lacks the warning, to avoid division by zero; this biases amplification scores for non-exposed users and distorts algorithm rankings.
  • Amplification evaluation horizon k = 100
    Headline amplification results (Figures 4-5) use k=100; accuracy uses k=10/50/100. The choice of horizon can change the relative ranking of algorithms.
  • SVD and ALS hyperparameters = included code
    SVD (factors, epochs, learning rate, regularization) and ALS (factors, regularization, positive-sample weight) tuned to RMSE and recall at k=50 respectively; values are in the included code. Not tuned to the amplification outcome, so not circular, but relevant to accuracy claims.
  • AO3 evaluation subset size = 10% of works
    Accuracy metrics for AO3 are computed on a 10% of-works subset (3.9M interactions) due to computational constraints; representativeness of this subset is not analyzed.
axioms (6)
  • domain assumption DDD community vote totals are valid, transferable sensitivity labels for the matched movies
    Section 3.1: labels binarized at 75% vote share without validation; all permutation and amplification results inherit errors from this rule.
  • domain assumption Webis Trigger Warning Corpus labels correctly describe the AO3 works in the collected subset
    Section 3.2: labels inherited from Wiegmann et al. [51] and interactions collected for only ~30% of works, without representativeness analysis.
  • domain assumption MovieLens IMDb/TMDB identifiers match DDD entries with negligible error
    Section 3.1: matching procedure described but unvalidated; the 52% match rate means the dataset covers under half of MovieLens movies.
  • domain assumption Public kudos are a valid implicit interaction signal despite unobserved guest kudos
    Section 3.2: the paper notes identifiable kudos do not equal total kudos, yet interaction density and all AO3 user-level analyses rely on public kudos alone.
  • standard math Permutation test exchangeability and matrix factorization convergence
    Section 4 and Section 5: standard statistical and algorithmic machinery with no special assumptions invoked.
  • ad hoc to paper Inserting a hallucinated warning item into a user's history does not distort amplification
    Section 5.2: this is an acknowledged hack, not a justified modeling assumption; it changes what the metric measures for users with no warning history.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Datasets for Navigating Sensitive Topics in Recommendation Systems." pith.science (2026). https://pith.science/paper/MZMA6SYA

@misc{pith2026250907269,
  author       = {Pith},
  title        = {Pith review of: Datasets for Navigating Sensitive Topics in Recommendation Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MZMA6SYA}},
  note         = {Machine review of arXiv:2509.07269}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Personalized AI systems, from recommendation systems to chatbots, are a prevalent method for distributing content to users based on their learned preferences. However, there is growing concern about the adverse effects of these systems, including their potential tendency to expose users to sensitive or harmful material, negatively impacting overall well-being. To address this concern quantitatively, it is necessary to create datasets with relevant sensitivity labels for content, enabling researchers to evaluate personalized systems beyond mere engagement metrics. To this end, we introduce two novel datasets that include a taxonomy of sensitivity labels alongside user-content ratings: one that integrates MovieLens rating data with content warnings from the Does the Dog Die? community ratings website, and another that combines fan-fiction interaction data and user-generated warnings from Archive of Our Own.

Figures

Figures reproduced from arXiv: 2509.07269 by Amelia Kovacs, Jerry Chee, Kimia Kazemian, Sarah Dean.

Figure 2
Figure 2. Figure 2: Distribution shift of average ratings or public kudos [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Per user average ratings or public kudos for works [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of amplification scores (ratio of warn [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Amplification of the ML-DDD "Is there blood/gore" [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

60 extracted references · 51 canonical work pages · 7 internal anchors

  1. [1]

    Himan Abdollahpouri, Masoud Mansoury, Robin Burke, and Bamshad Mobasher

  2. [2]

    Nil-Jana Akpinar, Liu Leqi, Dylan Hadfield-Menell, and Zachary Lipton. 2022. Counterfactual Metrics for Auditing Black-Box Recommender Systems for Ethical Concerns. InWorkshop on Responsible Decision Making in Dynamic Environments, International Conference on Machine Learning, ICML, Vol. 162

    work page 2022

  3. [3]

    Hind Almerekhi, Haewoon Kwak, Joni Salminen, and Bernard J Jansen. 2020. Are these comments triggering? predicting triggers of toxicity in online discussions. InProceedings of the web conference 2020. 3033–3040

    work page 2020

  4. [4]

    Chumki Basu, Haym Hirsh, and William W. Cohen. 1998. Recommendation as Classification: Using Social and Content-Based Information in Recommendation. InAAAI/IAAI. https://api.semanticscholar.org/CorpusID:14176567

    work page 1998

  5. [5]

    Paul Bouchaud. 2024. Algorithmic Amplification of Politics and Engagement Maximization on Social Media. InComplex Networks & Their Applications XII

    work page 2024

  6. [6]

    Jiawei Chen, Hande Dong, Xiang Wang, Fuli Feng, Meng Wang, and Xiangnan He. 2023. Bias and debias in recommender system: A survey and future directions. ACM Transactions on Information Systems41, 3 (2023), 1–39

    2023

  7. [7]

    Mihaela Curmei, Sarah Dean, and Benjamin Recht. 2021. Quantifying avail- ability and discovery in recommender systems via stochastic reachability. In International Conference on Machine Learning. PMLR, 2265–2275

    work page 2021

  8. [8]

    Sarah Dean, Sarah Rich, and Benjamin Recht. 2020. Recommendations and user agency: the reachability of collaboratively-filtered information. InProceedings of the 2020 conference on fairness, accountability, and transparency. 436–445

    work page 2020

  9. [9]

    Brianna Dym and Casey Fiesler. 2020. Ethical and Privacy Considerations for Research Using Online Fandom Data.Transformative works and cultures33 (2020)

    work page 2020

  10. [10]

    Michael D Ekstrand, Anubrata Das, Robin Burke, Fernando Diaz, et al . 2022. Fairness in information access systems.Foundations and Trends®in Information Retrieval16, 1-2 (2022), 1–177

    2022

  11. [11]

    Michael D Ekstrand, Mucun Tian, Ion Madrazo Azpiazu, Jennifer D Ekstrand, Oghenemaro Anuyah, David McNeill, and Maria Soledad Pera. 2018. All the cool kids, how do they fit in?: Popularity and demographic biases in recom- mender evaluation and effectiveness. InConference on fairness, accountability and transparency. PMLR, 172–186

    work page 2018

  12. [12]

    Michael D Ekstrand, Mucun Tian, Mohammed R Imran Kazi, Hoda Mehrpouyan, and Daniel Kluver. 2018. Exploring author gender in book rating and recom- mendation. InProceedings of the 12th ACM conference on recommender systems. 242–250

    work page 2018

  13. [13]

    Marc Faddoul, Guillaume Chaslot, and Hany Farid. 2020. A longitudinal analysis of YouTube’s promotion of conspiracy videos.preprint arXiv:2003.03318(2020)

    Pith/arXiv arXiv 2020

  14. [14]

    Miriam Fernández, Alejandro Bellogín, and Iván Cantador. 2021. Analysing the effect of recommendation algorithms on the amplification of misinformation. arXiv preprint arXiv:2103.14748(2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  15. [15]

    Frederickson. 2023. Implicit. benfred.github.io/implicit/

    work page 2023

  16. [16]

    Joshua Goodman. 2020. Making TikTok more accessible to people with photo- sensitive epilepsy. (November 2020)

    work page 2020

  17. [17]

    Rebecca Gormann and Stuart Armstrong. 2022. The dangers in algorithms learning humans’ values and irrationalities.arXiv preprint arXiv:2202.13985 (2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  18. [18]

    Wenshuo Guo, Karl Krauth, Michael Jordan, and Nikhil Garg. 2021. The stereo- typing problem in collaboratively filtered recommender systems. InProceedings of the 1st ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization. 1–10

    work page 2021

  19. [19]

    Maxwell Harper and Joseph A

    F. Maxwell Harper and Joseph A. Konstan. 2015. The MovieLens Datasets: History and Context.ACM Trans. Interact. Intell. Syst.5, 4 (2015)

    work page 2015

  20. [20]

    Herlocker, Joseph A

    Jonathan L. Herlocker, Joseph A. Konstan, Loren G. Terveen, and John T. Riedl

  21. [21]

    Yifan Hu, Yehuda Koren, and Chris Volinsky. 2008. Collaborative Filtering for Implicit Feedback Datasets. InEighth IEEE ICDM. 263–272

    work page 2008

  22. [22]

    Nicolas Hug. 2020. Surprise: A Python library for recommender systems.Journal of Open Source Software5, 52 (2020), 2174. doi:10.21105/joss.02174

    work page doi:10.21105/joss.02174 2020

  23. [23]

    Ferenc Huszár, Sofia Ira Ktena, Conor O’Brien, Luca Belli, Andrew Schlaikjer, and Moritz Hardt. 2021. Algorithmic Amplification of Politics on Twitter.CoRR abs/2110.11010 (2021). arXiv:2110.11010 https://arxiv.org/abs/2110.11010

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  24. [24]

    Instagram. 2021. Introducing Sensitive Content Control. (July 2021)

    work page 2021

  25. [25]

    Dietmar Jannach. 2022. Multi-Objective Recommender Systems: Survey and Challenges.arXiv preprint arXiv:2210.10309(2022)

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  26. [26]

    Benjamin Kille and Andreas Lommatzsch. 2019. Defining a Meaningful Baseline for News Recommender Systems. InINRA@RecSys

    work page 2019

  27. [27]

    Mark Ledwich, Anna Zaitsev, and Anton Laukemper. 2022. Radical bubbles on YouTube? Revisiting algorithmic extremism with personalised recommendations. First Monday(2022)

    work page 2022

  28. [28]

    Simon A Levin, Helen V Milner, and Charles Perrings. 2021. The dynamics of political polarization. e2116950118 pages

    work page 2021

  29. [29]

    Shao Yi Liaw, Fan Huang, Fabricio Benevenuto, Haewoon Kwak, and Jisun An

    work page

  30. [30]

    Kun Lin, Nasim Sonboli, Bamshad Mobasher, and Robin Burke. 2019. Crank up the volume: preference bias amplification in collaborative recommendation. arXiv preprint arXiv:1909.06362(2019)

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  31. [31]

    Liu Yi Lin, Jaime E Sidani, Ariel Shensa, Ana Radovic, Elizabeth Miller, Jason B Colditz, Beth L Hoffman, Leila M Giles, and Brian A Primack. 2016. Association between social media use and depression among US young adults.Depression and anxiety33, 4 (2016), 323–331

    work page 2016

  32. [32]

    2020.It’s almost impossible to avoid triggering content on TikTok

    Mashable. 2020.It’s almost impossible to avoid triggering content on TikTok. https://mashable.com/article/tiktok-algorithm-triggers

    work page 2020

  33. [33]

    Andriy Mnih and Russ R Salakhutdinov. 2007. Probabilistic Matrix Factorization. InAdvances in Neural Information Processing Systems, Vol. 20. , , Amelia Kovacs, Jerry Chee, Kimia Kazemian, and Sarah Dean

    work page 2007

  34. [34]

    Ioannis Mollas, Zoe Chrysopoulou, Stamatis Karlos, and Grigorios Tsoumakas

    work page

  35. [35]

    Dan S Nielsen and Ryan McConville. 2022. Mumin: A large-scale multilingual multimodal fact-checked misinformation social network dataset. InProceedings of the 45th international ACM SIGIR

    work page 2022

  36. [36]

    Nicholas J Restrepo, Lucia Illari, Rhys Leahy, Richard F Sear, Yonatan Lupu, and Neil F Johnson. 2021. How Social Media Machinery Pulled Mainstream Parenting Communities Closer to Extremes and their Misinformation during Covid-19. IEEE Access(2021)

    work page 2021

  37. [37]

    Manoel Horta Ribeiro, Raphael Ottoni, Robert West, Virgílio AF Almeida, and Wagner Meira Jr. 2020. Auditing radicalization pathways on YouTube. InProceed- ings of the 2020 conference on fairness, accountability, and transparency. 131–141

    work page 2020

  38. [38]

    Badrul Sarwar, George Karypis, Joseph Konstan, and John Riedl. 2000. Anal- ysis of recommendation algorithms for e-commerce. InProceedings of the 2nd ACM Conference on Electronic Commerce(Minneapolis, Minnesota, USA)(EC ’00). Association for Computing Machinery, New York, NY, USA, 158–167. doi:10.1145/352871.352887

    work page arXiv 2000

  39. [39]

    Dougal Shakespeare, Lorenzo Porcaro, Emilia Gómez, and Carlos Castillo

    work page

  40. [40]

    Renee Shelby, Shalaleh Rsimani, Kathryn Henne, AJung Moon, Negar Ros- tamzadeh, Paul Nicholar, N’Mah Yilla-Akbari, Jess Gallegos, Andrew SMart, Emilio Garcia, and Gurleen Virk. 2023. Sociotechnical Harms of Algorithmic Systems: Scoping a Taxonomy for Harm Reduction. InProceedings of the sixth AAAI/ACM Conference on Artificial Intelligence, Ethics, and Society

    work page 2023

  41. [41]

    Ashudeep Singh, Yoni Halpern, Nithum Thain, Konstantina Christakopoulou, EH Chi, Jilin Chen, and Alex Beutel. 2020. Building healthy recommendation sequences for everyone: A safe reinforcement learning approach. InFAccTRec Workshop

    work page 2020

  42. [42]

    Jessie J Smith, Lucia Jayne, and Robin Burke. 2022. Recommender Systems and Algorithmic Hate. InProceedings of the 16th ACM RecSys. 592–597

    work page 2022

  43. [43]

    Harald Steck. 2018. Calibrated recommendations. InProceedings of the 12th ACM conference on recommender systems. 154–162

    2018

  44. [44]

    Exploring artist gender bias in music recommendation.arXiv preprint arXiv:2009.01715(2020)

    work page internal anchor Pith review Pith/arXiv arXiv 2009

  45. [45]

    Özge Sürer, Robin Burke, and Edward C Malthouse. 2018. Multistakeholder recom- mendation with provider constraints. InProceedings of the 12th ACM Conference on Recommender Systems. 54–62

    work page 2018

  46. [46]

    Antonela Tommasel and Filippo Menczer. 2022. Do recommender systems make social media more susceptible to misinformation spreaders?. InProceedings of the 16th ACM Conference on Recommender Systems. 550–555

    work page 2022

  47. [47]

    Saúl Vargas and Pablo Castells. 2011. Rank and relevance in novelty and diversity metrics for recommender systems. InProceedings of the fifth ACM conference on Recommender systems. 109–116

    work page 2011

  48. [48]

    Bertie Vidgen and Leon Derczynski. 2020. Directions in abusive language training data, a systematic review: Garbage in, garbage out.Plos one15, 12 (2020)

    work page 2020

  49. [49]

    Wenlong Suna and Olfa Nasraouia. 2021. User Polarization Aware Matrix Factor- ization for Recommendation Systems. (2021)

    work page 2021

  50. [50]

    Joe Whittaker, Seán Looney, Alastair Reed, and Fabio Votta. 2021. Recommender systems and the amplification of extremist content.Internet Policy Review10, 2 (2021), 1–29

    work page 2021

  51. [51]

    Matti Wiegmann, Magdalena Wolska, Christopher Schröder, Ole Borchardt, Benno Stein, and Martin Potthast. 2023. Trigger warning assignment as a multi- label document classification problem. InProceedings of the 61st ACL

    work page 2023

  52. [52]

    Jobin Wilson, Santanu Chaudhury, Brejesh Lall, and Prateek Kapadia. 2014. Im- proving Collaborative Filtering based Recommenders using Topic Modelling. arXiv:1402.6238 [cs.IR]

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  53. [53]

    Ellery Wulczyn, Nithum Thain, and Lucas Dixon. 2017. Ex machina: Personal attacks seen at scale. InProceedings of the 26th international conference on world wide web. 1391–1399

    2017

  54. [54]

    2020.When targeted ads feel a little too targeted

    Vox. 2020.When targeted ads feel a little too targeted. https://www.vox.com/the- goods/2020/4/9/21204425/targeted-ads-fertility-eating-disorder-coronavirus

    work page 2020

  55. [55]

    Clear Yes

    Yong Zheng and David Xuejun Wang. 2022. A survey of recommender systems with multi-objective optimization.Neurocomputing474 (2022), 141–153. A Additional Information Table 3: The 36 warning labels used in the AO3 dataset and the number of works marked with the warning. Warning #Works pornography 174391 violence 42084 other-mental-health 40560 death 36123 ...

    work page 2022

  56. [59]

    Lin Xiao, Zhang Min, Zhang Yongfeng, Gu Zhaoquan, Liu Yiqun, and Ma Shaop- ing. 2017. Fairness-aware group recommendation with pareto-efficiency. In Proceedings of the eleventh ACM conference on recommender systems. 107–115

    work page 2017

  57. [2004]

    Evaluating collaborative filtering recommender systems.ACM Trans. Inf. Syst.22, 1 (jan 2004), 5–53. doi:10.1145/963770.963772

    work page arXiv 2004

  58. [2019]

    The impact of popularity bias on fairness and calibration in recommenda- tion.arXiv preprint arXiv:1910.05755(2019)

    Pith/arXiv arXiv 1910

  59. [2020]

    Ethos: an online hate speech detection dataset.arXiv:2006.08328(2020)

    Pith/arXiv arXiv 2006

  60. [2023]

    InProceedings of the International AAAI Conference on Web and Social Media, Vol

    YouNICon: YouTube’s CommuNIty of Conspiracy Videos. InProceedings of the International AAAI Conference on Web and Social Media, Vol. 17. 1102–1111

    work page

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.