Pith. sign in

REVIEW 3 major objections 5 minor 25 references

Exploration on Demand: From Algorithmic Control to User Empowerment

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a user-controlled exploration knob built on adaptive semantic clustering can meaningfully increase recommendation diversity, cutting intra-list similarity from 0.34 to 0.26 and raising unexpectedness to 0.73 on…

desk verdict An honest, modest diversity-layer paper whose abstract oversells 'balancing' and whose LLM A/B arm is not actually exploitative. read the letter →

arxiv 2507.21884 v2 pith:ACAIUSYC submitted 2025-07-29 cs.IR

Edoardo Bianchi This is my paper
classification cs.IR
keywords recommendersystemsexploration-exploitationtrade-offsemanticclusteringrecommendationdiversityuser-controlledexplorationLLM-basedevaluationunexpectednessMovieLens
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Recommender systems that only optimize for past clicks produce monotonous lists, and this paper asks whether users can be handed a control for how much novelty they want. It proposes an adaptive online clustering system that embeds movies with sentence-transformer vectors, groups them into semantic clusters with a dynamic similarity threshold, and sets aside a fraction of each recommendation list for clusters the user has rarely engaged with. On MovieLens data, turning on this exploration reduces intra-list similarity from 0.34 to 0.26 and raises unexpectedness from 0.67 to 0.73, and an LLM-based A/B test with 300 simulated users finds that 72.7% of long-history users prefer exploratory lists over exploitation-only ones. The paper's stated reading of its own results is modest: relevance metrics remain far below collaborative filtering and MMR, so the contribution is a controllable exploration layer that makes the relevance-diversity trade-off explicit rather than an accuracy-optimized recommender. What makes the claim worth caring about is that diversity becomes a user-adjustable property instead of an algorithmic side effect.

What carries the argument

The machinery is an adaptive online clustering algorithm combined with an explicit exploration budget. Each item becomes a 384-dimensional embedding of its metadata; a cosine-similarity threshold (initialized at 0.45 and periodically adjusted with silhouette scores) assigns items to clusters whose centroids are mean embeddings. Exploration then counts interactions per cluster over a recent window, ranks clusters by frequency, takes the top three as habitual, and fills $\lfloor \alpha k \rfloor$ list positions by uniform sampling from the remaining clusters, with default $\alpha=2/3$. This pair of mechanisms — cluster structure plus an $\alpha$-scaled sampling budget — is what turns generic diversity tuning into a user-specified allocation between familiar and unfamiliar semantic regions.

What would settle it

A human-subject A/B test on the $k=10$, $h=50$ condition would settle the preference claim: if long-history users do not pick exploratory lists at a rate clearly above fifty percent, the 72.7% result is an artifact of the LLM judge rather than a fact about users. A complementary check is whether held-out relevant items actually reside in the less-engaged clusters that exploration samples; the reported NDCG of 0.00 for the exploitation-only configuration already warns that those clusters may carry no relevance signal, which would undercut the interpretation of exploration as taste expansion.

Watch

Extended reading notes

Core claim

The paper's central discovery claim is that exploration can be organized at the semantic-cluster level and exposed as a user-controllable coefficient $\alpha$: items are embedded, clustered online by cosine similarity, and $\alpha$ decides what share of a list is sampled from less-engaged clusters. At $k=10$ with $h=50$ interactions of history, exploration lowers intra-list similarity from $0.34$ to $0.26$ and raises unexpectedness from $0.67$ to $0.73$; the same configuration records near-zero NDCG, Recall, and HitRate, which the authors take as the expected relevance-diversity trade-off. The LLM-based A/B test, run with 300 simulated users and the same MovieLens histories, shows that long-history users choose exploratory lists in 72.7% of cases at $k=5$, while short-history users lean exploitative. The paper therefore presents the system as a tunable discovery layer whose job is to widen exposure, with the cost in accuracy laid out in Table I rather than hidden.

Load-bearing premise

The central claim rests on the assumption that semantic clusters formed from movie titles, genres, tags, and descriptions correspond to real taste segments, so that less-engaged clusters are genuinely unexplored preferences rather than arbitrary partitions; if that fails, both the discovery benefit and the preference result lose their meaning.

Editorial extensions

If this is right

  • The exploration layer can be attached to an accuracy-oriented recommender without retraining it, because it reroutes list positions at the cluster level.
  • Interaction history length becomes a usable signal: exploration is preferred only after enough history exists, so a production system could scale $\alpha$ with history.
  • The sensitivity analysis over $\alpha$ gives a monotone control knob: raising $\alpha$ consistently lowers intra-list similarity and raises unexpectedness, allowing platforms to set an explicit diversity budget.
  • The same cluster representation serves both cold-start initialization and later exploration, so one semantic map supports onboarding and discovery.
  • The reported NDCG, Recall, and HitRate values quantify exactly how much accuracy is sacrificed, making the relevance-diversity trade-off auditable rather than implicit.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Combining cluster-level exploration with item-level ranking inside each selected cluster (for example, ranking candidates by collaborative-filtering scores after the cluster is chosen) would likely preserve most of the ILS and unexpectedness gains while recovering a substantial part of the lost relevance, because the near-zero NDCG stems from uniform sampling within clusters.
  • The LLM-based A/B protocol could be validated against a small human panel; if agreement is high, it becomes a cheap standard instrument for testing diversity-oriented recommenders, extending well beyond this system.
  • A natural product extension, implied but not built by the paper, is an adaptive $\alpha$ that grows with interaction history, so that new users receive familiar content and exploration appears only once the system can identify genuinely under-explored clusters.
  • The $\alpha$ parameter could be exposed directly in the user interface as an "explore more / stay familiar" slider, making the paper's control mechanism a visible user affordance rather than an internal setting.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes an adaptive semantic-clustering recommender with a user-controlled exploration mechanism. Items are embedded with Sentence-BERT, clustered online with a dynamic cosine-similarity threshold, and recommendations are sampled uniformly from the user's most-engaged clusters, with an exploration budget that draws from less-engaged clusters. Experiments on MovieLens 32M (20k-item subset and full catalog) report intra-list similarity, unexpectedness, NDCG/Recall/HitRate, and an LLM-based A/B test with 300 simulated users. The authors claim that exploration reduces ILS from 0.34 to 0.26, increases unexpectedness to 0.73, and that 72.7% of long-history users prefer exploratory lists, while candidly acknowledging that the method is a controllable exploration layer rather than an accuracy-optimized stand-alone recommender.

Significance. If the narrow claim—a simple, tunable diversity layer—were fully supported, this would be a modest but useful contribution: the online clustering procedure is clearly described, the sensitivity analysis over the exploration coefficient is a positive feature, and the larger-catalogue validation strengthens the diversity findings. The authors also deserve credit for explicitly quantifying the relevance cost and for stating in Section VII that item-level ranking is weak and that the LLM-based A/B test lacks human validation. However, the headline claims in the abstract and conclusion—'effectively balances personalization and diversity' and the user-preference result—are not supported by the reported evidence, because the proposed method has near-zero measured relevance and the A/B comparison does not use a genuine exploitative baseline. As it stands, the significance is that of a diversity-only add-on, not a balanced recommender, and the paper needs substantial revision before the claimed user-preference result can be accepted.

major comments (3)
  1. [Abstract and Table I] The abstract's claim that the system 'effectively balances personalization and diversity' is contradicted by Table I. Across all configurations, both 'Ours (Exploration Off)' and 'Ours (Exploration On)' achieve NDCG=0.00, Recall=0.00, and HitRate at most 0.03, while Collaborative Filtering reaches NDCG between 0.31 and 0.39 and HitRate between 0.65 and 0.80. This is not a balance between personalization and diversity; it is a diversity-only mechanism with no measurable relevance signal. The abstract, Section V, and Section VIII should be revised to state unambiguously that the contribution is a diversity/exploration layer, and the phrase 'effectively balances' should be removed or replaced with a precise statement of the trade-off as measured.
  2. [Section V-B and Algorithm 3] The LLM-based A/B test does not compare exploration against exploitation. In Algorithm 3, when 'explore=false', the algorithm still samples k items uniformly from the user's top-3 clusters, with no item-level relevance ranking. Table I confirms the 'Exploration Off' arm has NDCG=0.00 and HitRate at most 0.03, so it is effectively a random draw from broad semantic clusters. The reported 72.7% preference at k=5, h=50 therefore compares two unranked random lists that differ mainly in how many items come from outside the top clusters; it cannot support the claim that users prefer exploration over a personalized exploitative list. The authors should add a true exploitation arm (e.g., an item-level relevance ranking inside the top clusters, or a CF/MMR baseline) or explicitly reframe the result as a comparison of two cluster-sampling strategies.
  3. [Section V-A and Table I] The headline diversity gains are largely mechanical. Because the exploration budget is defined as sampling k_exp = floor(alpha*k) items from clusters outside the user's top-m set, and ILS and unexpectedness are computed from the same embeddings and cluster assignments used to define those clusters, the reduction in ILS (0.34 to 0.26) and increase in unexpectedness (0.67 to 0.73) are guaranteed by construction rather than evidence of meaningful discovery. To substantiate the discovery claim, the authors should include a control that samples random items from outside the top clusters without using user history, or show that exploratory items have measurable held-out relevance or user-level appeal beyond the confounded A/B test.
minor comments (5)
  1. [Algorithm 1, line 24] The condition 'merge clusters with similarity > 0.9' is underspecified; the authors should state whether cluster similarity is centroid cosine similarity, minimum pairwise similarity, or some other aggregate.
  2. [Equation (2)] The symbol '\oplus' is used for string concatenation but is never defined; please define it or use explicit concatenation notation.
  3. [Section IV-A] The description 'sampled 20,000 items while preserving the original data distribution across genres and popularity levels' lacks reproducibility details such as the sampling seed and procedure; please provide them.
  4. [Section V-B] There is a typographical double period at the end of 'more appealing recommendation lists..'; the sentence should end with a single period.
  5. [Section IV-D] The A/B prompt does not mention randomization or counterbalancing of Set A and Set B order across the 300 simulated users; without position counterbalancing, position bias in LLM judgments cannot be excluded.

Circularity Check

2 steps flagged · score 6.0 of 10

Headline diversity/unexpectedness gains are built into Algorithm 3's sampling rule; LLM A/B preference compares two unranked cluster samples, so only the relevance trade-off against CF/MMR is independent evidence.

  1. self definitional [Algorithm 3 (Section III-E), Section IV-C metric definitions, Section V-A results]
    "if explore then identify remaining clusters: T_other ← T \ T_top; set exploration budget: k_exp ← ⌊αk⌋; sample k_exp items from T_other ... Enabling exploration reduces ILS and increases Unexpectedness relative to the exploitation-only version, showing that the system effectively expands recommendation lists beyond the user's most familiar semantic regions. Diversity and novelty are measured through ILS [7] and Unexpectedness [8]."

    ILS is the average pairwise item similarity in the recommended list, and clusters are constructed from the same sentence-transformer cosine similarities (Algorithm 1 uses 'find nearest centroid T_n using cosine similarity; if sim(e, T_n) > τ then add id to cluster C_n'). T_other is defined as every cluster outside the user's top-3 most-engaged clusters, i.e., clusters whose items are, by construction, dissimilar to the user's history clusters. Sampling k_exp = floor(α k) items from T_other deterministically injects cross-cluster items into the list, lowering average pairwise similarity relative to the all-top-cluster list. Unexpectedness is a history-dissimilarity measure; items from 'less-engaged clusters' are by definition dissimilar to the user's engaged history.

  2. other [Section IV-D (LLM A/B setup), Algorithm 3 lines 15-17, Table I rows 'Ours(Exploration Off)']
    "For each simulated user, we generated two recommendation sets: one with exploration disabled ("exploitation-only") and another with exploration enabled. ... set exploitation budget: k_explt ← k − |R|; sample k_explt items from top clusters; add sampled items to R. ... Ours(Exploration Off) 0.01 0.00 0.03 0.33 0.67 27.3; Ours(Exploration On) 0.00 0.00 0.01 0.27 0.72 72.7"

    The 'exploitation-only' arm is not exploitation: Algorithm 3 still samples uniformly from the user's top-3 clusters with no item-level relevance score, and Table I shows NDCG=0.00, Recall=0.00, HitRate≤0.03 for every 'Exploration Off' row. The 72.7% A/B preference is therefore a comparison between two unranked random lists that differ mainly in whether a fraction of items comes from outside the top clusters. Calling the baseline 'purely exploitative' in the abstract renames cluster-level uniform sampling as personalization; the preference result cannot support the claimed relevance-diversity trade-off or the conclusion that exploration 'translates into more appealing recommendation lists' beyond a generic preference for variety.

full rationale

The diversity and unexpectedness results are not empirically derived: the exploration rule samples from clusters outside the user's engaged top clusters, and ILS and Unexpectedness are computed from the same embedding-based similarities used to form those clusters, so the reported improvements (0.34→0.26 and 0.67→0.73) follow by construction. The A/B preference claim is weakened because the 'exploitation-only' condition is also an unranked random draw from broad clusters, as confirmed by its zero NDCG in Table I; the comparison therefore tests variety preference, not exploration-vs-exploitation. The paper is partially honest about this through its relevance results and Section VII limitations, and the CF/MMR benchmarks provide independent evidence of the accuracy trade-off. However, the central headline gains reduce to the algorithm's own definitions, so a score of 6 is warranted rather than a higher score that would ignore the independent relevance benchmarking.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The core of the paper is an engineering system built from standard components, so the ledger is dominated by hand-chosen parameters rather than invented entities. The three load-bearing domain assumptions: embedding similarity tracks preference (Section III-B2), clusters match taste segments (Section III-E), and the DeepSeek-V3 judge approximates human preference (Section IV-D). The first two are directly testable and are weakly supported by the paper's own relevance results (NDCG = 0.00 even with exploration off). No new particles, mediators, forces, or conserved quantities are invented; the semantic cluster is a standard construct.

free parameters (8)
  • alpha exploration coefficient = 2/3 default; sensitivity over {0.25, 0.50, 0.66, 0.75}
    Controls the exploration budget k_exp = floor(alpha * k); hand-chosen default with no principled selection. Algorithm 3, Section V-C.
  • initial clustering threshold tau_0 = 0.45
    Hand-chosen initialization of the cluster join threshold; determines cluster granularity. Section IV-B.
  • threshold adjustment multipliers and bounds = 0.95, 0.98, 1.02; bounds [0.3, 0.8]
    Hand-chosen adaptation schedule for dynamic thresholding. Algorithm 1.
  • silhouette score regime bounds = 0.1, 0.2, 0.4
    Hand-chosen quality thresholds that decide whether tau is reduced or increased. Algorithm 1.
  • number of top clusters m = 3
    Number of clusters treated as the user's taste region in both cold-start and personalized recommendation. Algorithms 2 and 3.
  • threshold update frequency f = 100 items
    Hand-chosen cadence for silhouette evaluation; affects cluster evolution. Algorithm 1.
  • silhouette sample size = 1000 items
    Random sample used for scalable silhouette computation; hand-chosen. Section IV-B.
  • LLM temperature = 0.4
    Hand-chosen setting for A/B preference judgments. Section IV-D.
assumptions (5)
  • domain assumption all-MiniLM-L6-v2 embeddings of concatenated title, genres, tags, and description capture semantic similarity that is meaningful for recommendation relevance and taste.
    Invoked in Section III-B2; if embeddings do not track preference-relevant content, the clusters carry no meaning.
  • domain assumption Semantic content clusters correspond to user taste segments, so less-engaged clusters are unexplored preference regions rather than artifact groups.
    Invoked in Section III-E and Algorithm 3; undermined by the paper's own exploitation-off NDCG = 0.00, which suggests top clusters carry no measurable relevance signal.
  • domain assumption LLM preference judgments (DeepSeek-V3 at temperature 0.4) approximate real user preferences for recommendation lists.
    Invoked in Section IV-D, relying on Refs. [11]-[14]; the paper itself concedes no human validation in Section VII.
  • domain assumption Held-out MovieLens ratings of at least 4.0 define ground-truth relevance.
    Invoked in Section IV-C; standard practice, but the binary threshold discards relative preference information in the rating scale.
  • standard math Cosine similarity on embedding vectors, arithmetic-mean centroids, and silhouette scores behave as standard background results.
    Used throughout Section III-C; standard and uncontroversial.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploration on Demand: From Algorithmic Control to User Empowerment." pith.science (2026). https://pith.science/paper/ACAIUSYC

@misc{pith2026250721884,
  author       = {Pith},
  title        = {Pith review of: Exploration on Demand: From Algorithmic Control to User Empowerment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ACAIUSYC}},
  note         = {Machine review of arXiv:2507.21884}
}
read the original abstract

Recommender systems often struggle with over-specialization, which severely limits users' exposure to diverse content and creates filter bubbles that reduce serendipitous discovery. To address this fundamental limitation, this paper introduces an adaptive clustering framework with user-controlled exploration that effectively balances personalization and diversity in movie recommendations. Our approach leverages sentence-transformer embeddings to group items into semantically coherent clusters through an online algorithm with dynamic thresholding, thereby creating a structured representation of the content space. Building upon this clustering foundation, we propose a novel exploration mechanism that empowers users to control recommendation diversity by strategically sampling from less-engaged clusters, thus expanding their content horizons while explicitly exposing the relevance-diversity trade-off. Experiments on the MovieLens dataset demonstrate the system's effectiveness, showing that exploration significantly reduces intra-list similarity from 0.34 to 0.26 while simultaneously increasing unexpectedness to 0.73. Furthermore, our Large Language Model-based A/B testing methodology, conducted with 300 simulated users, reveals that 72.7% of long-term users prefer exploratory recommendations over purely exploitative ones. Additional relevance metrics, including NDCG@k, Recall@k, and HitRate@k, reveal the expected relevance-diversity trade-off against CF and MMR baselines, positioning the method as a controllable exploration layer for promoting meaningful content discovery.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 15 canonical work pages

  1. [1]

    Trends in content-based recommendation

    M. C. e. a. Lops P., Jannach D., “Trends in content-based recommendation.” 2019. [Online]. Available: https://doi.org/10.1007/ s11257-019-09231-w

    work page 2019

  2. [2]

    Leveraging deep visual features for content-based movie recommender systems,

    R. J. R. Filho, J. Wehrmann, and R. C. Barros, “Leveraging deep visual features for content-based movie recommender systems,” in 2017 International Joint Conference on Neural Networks (IJCNN), 2017, pp. 604–611

    work page 2017

  3. [3]

    Overview on nlp techniques for content-based recommender systems for books,

    M. Berbatova, “Overview on nlp techniques for content-based recommender systems for books,” 2019. [Online]. Available: https://doi.org/10.26615/issn.2603-2821.2019 00955

    arXiv 2019

  4. [4]

    Novel Item Recommendation by User Profile Partitioning,

    M. Zhang and N. Hurley, “Novel Item Recommendation by User Profile Partitioning,” inWeb Intelligence and Intelligent Agent Technologies, 2009. WI-IAT '09. IEEE/WIC/ACM International Joint Conferences on, vol. 1. Los Alamitos, CA, USA: IET, 2009, pp. 508–515. [Online]. Available: http://dx.doi.org/10.1109/wi-iat.2009.85

    work page doi:10.1109/wi-iat.2009.85 2009

  5. [5]

    Online k-means Clustering

    V . Cohen-Addadet al., “Online k-means clustering,” 2019. [Online]. Available: https://doi.org/10.48550/arXiv.1909.06861

    work page Pith review arXiv doi:10.48550/arxiv.1909.06861 2019

  6. [6]

    Online clustering with experts,

    A. Choromanska and C. Monteleoni, “Online clustering with experts,” inProceedings of the Fifteenth International Conference on Artificial Intelligence and Statistics, ser. Proceedings of Machine Learning Research, N. D. Lawrence and M. Girolami, Eds., vol. 22. La Palma, Canary Islands: PMLR, 21–23 Apr 2012, pp. 227–235. [Online]. Available: https://procee...

    work page 2012

  7. [7]

    Improving recommendation lists through topic diversification,

    C.-N. Ziegleret al., “Improving recommendation lists through topic diversification,” inProceedings of the 14th International Conference on World Wide Web, ser. WWW ’05. New York, NY , USA: Association for Computing Machinery, 2005, p. 22–32. [Online]. Available: https://doi.org/10.1145/1060745.1060754

    arXiv 2005

  8. [8]

    On unexpectedness in recommender systems: Or how to better expect the unexpected,

    P. Adamopoulos and A. Tuzhilin, “On unexpectedness in recommender systems: Or how to better expect the unexpected,”ACM Trans. Intell. Syst. Technol., vol. 5, no. 4, dec 2014. [Online]. Available: https://doi.org/10.1145/2559952

    work page doi:10.1145/2559952 2014

Show all 25 references
  1. [9]

    C. C. Aggarwal,Recommender Systems: The Textbook. Springer, 2016

    2016

  2. [10]

    B. S. Francesco Ricci, Lior Rokach,Recommender Systems Handbook, 2nd ed. Springer, 2015

    2015

  3. [11]

    Simuser: Simulating user behavior with large language models for recommender system evaluation,

    N. Bougie and N. Watanabe, “Simuser: Simulating user behavior with large language models for recommender system evaluation,”

  4. [12]

    Llm-powered user simulator for recommender system,

    Z. Zhanget al., “Llm-powered user simulator for recommender system,” 2024. [Online]. Available: https://arxiv.org/abs/2412.16984

    2024 arXiv

  5. [13]

    Can large language model agents simulate human trust behavior?

    C. Xieet al., “Can large language model agents simulate human trust behavior?” 2024. [Online]. Available: https://arxiv.org/abs/2402.04559

    2024 arXiv

  6. [14]

    Agent hospital: A simulacrum of hospital with evolvable medical agents,

    J. Liet al., “Agent hospital: A simulacrum of hospital with evolvable medical agents,” 2025. [Online]. Available: https://arxiv.org/abs/2405. 02957

    2025

  7. [15]

    Sentence-bert: Sentence embeddings using siamese bert-networks,

    N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert-networks,” inProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 11 2019. [Online]. Available: https://arxiv.org/ab...

    2019 arXiv

  8. [16]

    Silhouettes: A graphical aid to the interpretation and validation of cluster analysis,

    P. J. Rousseeuw, “Silhouettes: A graphical aid to the interpretation and validation of cluster analysis,”Journal of Computational and Applied Mathematics, vol. 20, pp. 53–65, 1987. [Online]. Available: https://www.sciencedirect.com/science/article/pii/0377042787901257

    1987

  9. [17]

    The movielens datasets: History and context,

    F. M. Harper and J. A. Konstan, “The movielens datasets: History and context,”ACM Trans. Interact. Intell. Syst., vol. 5, no. 4, Dec

  10. [18]

    Cumulated gain-based evaluation of ir techniques,

    K. J ¨arvelin and J. Kek ¨al¨ainen, “Cumulated gain-based evaluation of ir techniques,”ACM Trans. Inf. Syst., vol. 20, no. 4, p. 422–446, Oct

  11. [19]

    Quality metrics in recommender systems: Do we calculate metrics consistently?

    Y .-M. Tamm, R. Damdinov, and A. Vasilev, “Quality metrics in recommender systems: Do we calculate metrics consistently?” Proceedings of the 15th ACM Conference on Recommender Systems,

  12. [20]

    Deepseek-v3 technical report,

    DeepSeek-AIet al., “Deepseek-v3 technical report,” 2025

    2025

  13. [21]

    The use of mmr, diversity-based reranking for reordering documents and producing summaries,

    J. Carbonell and J. Goldstein, “The use of mmr, diversity-based reranking for reordering documents and producing summaries,” in Proceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, ser. SIGIR ’98. New York, NY ...

    1998 doi

  14. [2002]

    Available: https://doi.org/10.1145/582415.582418

    [Online]. Available: https://doi.org/10.1145/582415.582418

  15. [2015]

    Available: https://doi.org/10.1145/2827872

    [Online]. Available: https://doi.org/10.1145/2827872

    doi

  16. [2021]

    Available: https://api.semanticscholar.org/CorpusID: 237495036

    [Online]. Available: https://api.semanticscholar.org/CorpusID: 237495036

  17. [2025]

    Available: https://arxiv.org/abs/2504.12722

    [Online]. Available: https://arxiv.org/abs/2504.12722

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.