Pith. sign in

REVIEW 4 major objections 7 minor 45 references

A Bayesian Choice Model for Eliminating Feedback Loops

T0 review · 4 major / 7 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read The paper presents the Dirichlet-Luce model, a Bayesian choice model in which the posterior marginal of an option that was never presented remains exactly equal to its prior, eliminating the negative bias that drives feedback loops.

desk verdict Sound fairness lemma in a Luce-based choice model, but the title oversells and the sample-complexity claim needs proof; worth refereeing. read the letter →

arxiv 1908.05640 v2 pith:ZYXUTCRP submitted 2019-08-15 stat.ML cs.LG

Gökhan Çapan , Ilker Gündoğdu , Ali Caner Türkmen , Çağr{i} Sofuoğlu , Ali Taylan Cemgil This is my paper · ORCID
classification stat.MLcs.LG
keywords BayesianchoicemodelLuceaxiomfeedbackloopsrecommendersystemsThompsonsamplingcold-startfairnesssequentialMonteCarlobias
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

The paper introduces a Bayesian choice model, the Dirichlet-Luce model, for interactive systems in which users choose from small, systematically chosen subsets of a large option set. Its central claim is fairness: for any option that has never been presented, the posterior marginal equals the prior marginal exactly, so the model never penalizes unexplored or cold-started alternatives. Because this correction is built into the likelihood rather than added as a penalty, the model also yields accurate preference estimates from relatively few pairwise-type observations and supports a Thompson-sampling presentation mechanism with low regret. If the model is right, feedback-loop bias in recommender-style systems can be removed at the level of the choice model itself.

What carries the argument

The load-bearing object is the generalized Dirichlet family $$p(\$\theta$ \mid \$\alpha$, \$\beta$) \propto \prod_k \$theta_k^{{\alpha_k - 1}}$ \prod_{C \in \mathcal{C}} \left(\sum_{\kappa \in C} \theta_\kappa\right)^{-\$\beta$(C)},$$ with consistency $\sum_C \beta(C) = \sum_k \alpha_k$; it reduces to the ordinary Dirichlet when $\beta$ is placed only on the full set $[K]$. This family carries the argument because the denominator factors for each presentation $C$ are the mechanism that removes negative bias: an option absent from $C$ never appears in the sums, so its marginal stays untouched. The same family makes sequential Monte Carlo weight updates $O(L)$ and supports the Thompson-sampling presentation of top $L$ sampled options.

What would settle it

Run a direct numerical check of the invariance claim: present only options 1 and 2 many times, keep option 3 never presented, and verify at every step that the posterior marginal of $\theta_3$ remains exactly the prior Beta marginal; any deviation disproves Lemma 3. More behaviorally, simulate a user with a Condorcet cycle (A preferred to B, B to C, C to A) and check whether the Thompson-sampling presentation mechanism's cumulative regret grows linearly rather than sublinearly, indicating that the single-transitive-preference assumption cannot capture the feedback.

Watch

Extended reading notes

Core claim

The central discovery is the Dirichlet-Luce posterior $$p(\$\theta$ \mid k_{1:T}, C_{1:T}, \$\alpha$, \$\beta$) \propto \prod_{k} \$theta_k^{{\alpha_k + y_k - 1}}$ \prod_{C \in \mathcal{C}} \left(\sum_{\kappa \in C} \theta_\kappa\right)^{-\mu(C)-\$\beta$(C)},$$ which accounts for limited exposure by putting each presentation's sum into the denominator. The paper proves (Lemma 3) that when an option is absent from every presentation, its posterior marginal is exactly its prior Beta marginal, independent of all observed choices; this is what makes the model fair to unpresented and cold-start options. The same construction gives a family whose normalization is Carlson's $R$ function, for which exact inference is intractable but Monte Carlo sampling is efficient. The paper argues these properties translate into an online presentation algorithm: sample a preference vector from the posterior, present its top $L$ options, observe the chosen item, and update.

Load-bearing premise

The load-bearing premise is that each user is governed by one fixed preference vector satisfying Luce's choice axiom, so all choices are stochastically transitive; if real users are intransitive or context-dependent, the model treats their conflicting choices as random noise and the fairness and regret guarantees no longer describe their behavior.

Editorial extensions

If this is right

  • Never-presented options, including newly added ones, keep their prior preference distribution, so cold-start items are neither downranked nor upranked by choices involving other options.
  • Preference recovery is sample-efficient: under a transitive preference assumption, accurate estimates require on the order of $O(K \log K)$ unique pairwise presentations, matching stochastic ranking algorithms.
  • The Thompson-sampling presentation mechanism achieves lower cumulative regret than TopRank for $L$-wise presentations and than Double Thompson Sampling for pairwise duels in simulated Plackett-Luce and cyclic preference settings.
  • The model can be reused as a building block in recommender systems and other interactive personalization pipelines to avoid filter-bubble effects.

Reading between the lines

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

  • A testable extension is a mixture of Dirichlet-Luce components: the paper notes that cyclical or multi-modal preferences can be captured as mixtures, which would let the same machinery model users who switch between transitive preference states.
  • The invariance property suggests a broader design principle: if a choice model's likelihood factors over options in the numerator and controls each presented subset's denominator, cold-start neutrality follows automatically; this could be transplanted to collaborative filtering.
  • Because the model is misspecified under intransitive preferences, real deployments might need a diagnostic that checks whether observed choices are consistent with a single transitive preference vector; the paper's treatment of conflicting choices implies such a diagnostic could be based on how much posterior mass is placed on random draws versus stable structure.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper introduces the Dirichlet-Luce model, a Bayesian choice model in which the likelihood is a restricted multinomial conditioned on presented subsets and the prior is a generalized Dirichlet density with presentation pseudo-counts. The authors show that the posterior marginal of an option that was never presented equals its prior marginal (Lemma 3, Appendix A.1.3), and they argue this makes the model fair to cold-start and underrepresented options. They then propose a Thompson-sampling presentation mechanism with sequential Monte Carlo inference, and report synthetic experiments showing lower cumulative regret than TopRank (for L-wise presentations) and Double Thompson Sampling (for pairwise duels), including one cyclic-preference scenario. The paper claims that the model is practical because accurate preference estimates are obtained from O(K log K) unique presentations, and that it eliminates feedback loops.

Significance. If the advertised claims held in their strongest form, the paper would be a valuable contribution: the invariance property in Lemma 3 is a genuine and non-obvious consequence of the Dirichlet-Luce posterior, and the SMC-based Thompson sampling mechanism is a sensible way to operationalize it. The proof of Lemma 3 is correct, and the model provides a clean Bayesian treatment of choice under limited exposure. The significance is reduced, however, by the gap between what is proved and what is advertised: the O(K log K) sample-complexity claim is a heuristic supported only by an analogy and a synthetic experiment; the regret results are purely empirical and lack theoretical bounds; and the 'eliminating feedback loops' language in the title and abstract is stronger than the formal results, which are conditional on a static Luce-preference model.

major comments (4)
  1. [Abstract; Section 3, 'Preference Learning'] The abstract and the bullet list claim that 'Dirichlet-Luce posterior inference achieves pairwise preference aggregation upon collecting statistics for O(K log K) unique pairs.' No sample-complexity theorem is stated or proved. The argument in Section 3 is an analogy with stochastic ranking algorithms, and the evidence is the empirical demonstration in Figure 1 that θ* can be recovered from Merge-Rank-generated comparisons. This is a heuristic observation about a specific active-sampling procedure, not a property of the Dirichlet-Luce model in general. Since the 'practical' claim in the abstract rests on this point, please either provide a formal sample-complexity bound (e.g., PAC-style posterior concentration) or explicitly label the O(K log K) statement as an empirical finding and remove it from the bullet list of model properties.
  2. [Title; Abstract; Section 4] The title and abstract claim that the model eliminates feedback loops and that the resulting system is 'free of feedback loops.' What is formally proved (Lemma 3, Appendix A.1.3) is marginal invariance for never-presented options under a single static preference vector satisfying Luce's axiom. Section 4 itself concedes that presented but rarely chosen options can still suffer transitivity-induced negative bias ('the mean preference estimate of θ2 will still decrease'), and the proposed Thompson-sampling remedy is heuristic. No regret bound, bias bound, or convergence guarantee is provided for the online presentation mechanism. The contribution should be reframed as reducing negative bias toward unpresented options under the Luce/static-preference regime, and the feedback-loop language should be reserved for what is actually demonstrated.
  3. [Section 3, 'Conflicting Choices'] The paper acknowledges that the likelihood depends only on the sufficient statistics μ(C) and y_k, and therefore discards all associations between specific choices and specific presentations. Consequently, Lemma 3's 'fairness' statement is about the posterior of a latent parameter θ in a model that is misspecified for context-dependent or intransitive users; it does not by itself establish fair treatment of cold-start items in real interactive systems. The abstract's unqualified 'fair' claim and the 'Fair Cold-Start' discussion in Section 3 are only justified inside the stochastic-transitivity regime. Please state this condition explicitly at the point of the claims, and address what the cyclic experiment in Figure 5c does and does not show (average-case improvement with high variance, no formal degradation bound).
  4. [Section 6, 'L-wise Presentations'] The comparison with TopRank does not specify how TopRank's interaction protocol is aligned with the paper's presentation-and-feedback model. The regret definition in Section 6 measures the top-N options 'included in the presentation,' but TopRank is a ranking algorithm; if its feedback model or evaluation differs from the restricted-multinomial choice model used for the simulated user, the cumulative-regret comparison may not be apples-to-apples. Please state the exact protocol: how TopRank's ranking is converted to a presentation, which feedback it receives, and how the learner's subset reward is computed in each round. This is needed to validate the central empirical claim of low regret in the L-wise setting.
minor comments (7)
  1. [Section 2] The notation uses C both for the set of all non-empty subsets and for a particular presentation C_t; please use a different symbol (e.g., a calligraphic C) to avoid confusion.
  2. [Section 3, 'Preference Learning'] The statement that 'the posterior is log-concave with respect to log-preferences' is given without proof or reference; please provide a brief argument or a citation.
  3. [Section 3, Eq. (2)] The consistency condition ∑_C β(C) = ∑_k α_k is introduced as a requirement; please justify why it is needed beyond the reduction to the Dirichlet prior at β=β0, and discuss what happens when it is violated.
  4. [Section 6, Figure 5 caption] The caption refers to the proposed algorithm as 'JTS,' which is not defined anywhere in the paper; please name it consistently (e.g., 'Dirichlet-Luce Thompson sampling').
  5. [Section 6; Appendix A.4] The number of simulation runs and key hyperparameter settings for the proposed algorithm (number of particles N, resampling threshold, Metropolis proposal parameters) are not reported for Figures 4 and 5; please add these details to the supplementary material.
  6. [Appendix A.3, Algorithm 2] The move step proposes θ̂_j uniformly on (0, r), which can have very low acceptance probability for peaked posteriors; please report the empirically observed acceptance rate or justify the choice of proposal.
  7. [Section 3, 'Conflicting Choices'] The sentence 'Cyclical choice behavior can be modeled as a mixture of preferences θ' is presented as future work, but the paper does not discuss how the posterior invariance property would extend to a mixture; please add a sentence clarifying that the current formal results do not cover the mixture case.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central fairness property is proved from the model's likelihood and prior, and the bandit comparisons use external baselines.

full rationale

The paper's fairness claim (Lemma 3, Appendix A.1.3) is derived from the Dirichlet-Luce posterior (Eq. 3). For an option l that never appears in any presentation, the likelihood factorizes so that θ_l enters only through the Dirichlet prior; the proof integrates out the remaining coordinates and shows the marginal is Beta(α_1, Σ_{j≠1} α_j), identical to the prior. This is a mathematical consequence of the model, not an input defined in terms of the target claim. The bandit results compare Dirichlet-Luce Thompson sampling against external baselines (TopRank, DTS) on simulated feedback, with no parameter fitted to the metric being reported. The sample-complexity statement O(K log K) is supported by analogy to stochastic ranking and by the Merge-Rank experiment (Fig. 1); it is an empirical overstatement rather than a circular reduction, since the experiment does not fit a parameter and then predict it. No load-bearing self-citations are present; references to Davidson-Solomon, Dickey, Jiang et al., and Hankin are prior-art attributions. Consequently, no step of the derivation reduces by construction to its inputs.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The model relies on Luce's axiom and a static, transitive preference vector; the fairness property is a consequence of the Dirichlet prior and the restricted multinomial likelihood. The prior hyperparameters are chosen, not fitted. No new physical entities are introduced.

free parameters (2)
  • α (Dirichlet concentration parameters) = not specified; assumed flat prior (α=1) in experiments
    Prior hyperparameters chosen by hand; they scale the prior pseudo-counts. The fairness property holds for any α, so the central derivation does not depend on a fitted value.
  • β (presentation pseudo-counts) = β0([K])=sum_k α_k, 0 otherwise
    Set to β0 so the prior is Dirichlet; a modeling choice, not fitted.
assumptions (4)
  • domain assumption User choices are generated by a static preference vector θ under Luce's choice axiom (Plackett-Luce restricted multinomial).
    Section 2 defines the choice model as p(k|C) = θ_k / sum_{κ in C} θ_κ. This is load-bearing: the posterior and the fairness property are derived from this form.
  • domain assumption Stochastic transitivity: a single θ implies preferences are transitive in expectation.
    Section 3 relies on transitivity to argue O(K log K) sample complexity via analogy to sorting. The paper acknowledges intransitive users as a limitation.
  • ad hoc to paper The consistency condition sum_C β(C) = sum_k α_k.
    Introduced in Section 3 to ensure prior parameters act as pseudo-counts of a contingency table; needed for the model to reduce to Dirichlet when β=β0.
  • standard math Monte Carlo sampling (SMC, Metropolis-within-Gibbs) approximates the posterior.
    Used throughout Section 4 and Appendix A.3; for large K, the posterior is intractable, so inference is approximate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Bayesian Choice Model for Eliminating Feedback Loops." pith.science (2026). https://pith.science/paper/ZYXUTCRP

@misc{pith2026190805640,
  author       = {Pith},
  title        = {Pith review of: A Bayesian Choice Model for Eliminating Feedback Loops},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZYXUTCRP}},
  note         = {Machine review of arXiv:1908.05640}
}
read the original abstract

Self-reinforcing feedback loops in personalization systems are typically caused by users choosing from a limited set of alternatives presented systematically based on previous choices. We propose a Bayesian choice model built on Luce axioms that explicitly accounts for users' limited exposure to alternatives. Our model is fair---it does not impose negative bias towards unpresented alternatives, and practical---preference estimates are accurately inferred upon observing a small number of interactions. It also allows efficient sampling, leading to a straightforward online presentation mechanism based on Thompson sampling. Our approach achieves low regret in learning to present upon exploration of only a small fraction of possible presentations. The proposed structure can be reused as a building block in interactive systems, e.g., recommender systems, free of feedback loops.

Figures

Figures reproduced from arXiv: 1908.05640 by the authors.

Figure 1
Figure 1. Estimated (from 1000 samples) posterior mean (over 50 runs) conditioned on an actively selected data set (by Merge-Rank with bias and confidence pa￾rameters  = 0.05 and δ = 0.1) of presentations due to transitivity and with simulated choices. θ ∗ is ordered, and E[θ] estimates are conformably permuted for visu￾alization. Shaded region denotes the standard deviation. In this light, we explore whether the Dirichlet-L… view at source ↗
Figure 2
Figure 2. Evolution of the exact posterior for T = 6 observations. Cyclic preferences (b) are treated as draws (a). (0, 0, 1) (1, 0, 0) (0, 1, 0) 1 0 1 0 1 0 1 2 3 (a) Prior (0, 0, 1) (1, 0, 0) (0, 1, 0) 1 0 1 0 1 0 1 2 3 (b) p(θ | k1:T , C1:T , α, β0) (0, 0, 1) (1, 0, 0) (0, 1, 0) 1 0 1 0 1 0 1 2 3 (c) Dirichlet posterior p(θ | k1:T , α), C1:T is ignored [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Prior choice probabilities, along with the posterior where options [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Average cumulative regret at top-N alternatives included in presentations in online learning to rank scenario after round T = 10000. Error bars denote the standard deviation. TopRank hyperparameter δ was optimized in a held-out experiment. Learning from Pairwise Prefer…
Figure 5
Figure 5. Figure 5: Average cumulative weak dueling regret (lower is better) in dueling bandits setup for [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Average cumulative regret (over 10 runs) at top-2 options (out of 100) included in presenta [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Total absolute deviation and Kullback-Leibler divergence of the estimated [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Posterior distribution of preferences under Dirichlet-Luce and Dirichlet-Multinomial [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: b Finally for online learning to rank experiments with presentation size L = 5, we used the θ ∗ in Figure 9c 1 5 10 15 20 25 30 35 40 45 50 k 0.00 0.05 0.10 0.15 0.20 0.25 *k Sparse * Dense * (a) θ ∗ in Dueling Bandits Experiments 1 5 10 15 20 25 30 35 40 45 50 55 60 6…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 45 canonical work pages

  1. [1]

    Random U tility T heory for S ocial C hoice

    Hossein Azari, David Parks, and Lirong Xia. Random U tility T heory for S ocial C hoice. In Advances in Neural Information Processing Systems, pages 126--134, 2012

    work page 2012

  2. [2]

    Rank A nalysis of I ncomplete B lock D esigns: I

    Ralph Allan Bradley and Milton E Terry. Rank A nalysis of I ncomplete B lock D esigns: I. the M ethod of P aired C omparisons. Biometrika, 39 0 (3/4): 0 324--345, 1952

    work page 1952

  3. [3]

    Top-k S election based on A daptive S ampling of N oisy P references

    R \'o bert Busa-Fekete, Balazs Szorenyi, Weiwei Cheng, Paul Weng, and Eyke H \"u llermeier. Top-k S election based on A daptive S ampling of N oisy P references. In International Conference on Machine Learning, pages 1094--1102, 2013

    work page 2013

  4. [4]

    u llermeier, and Bal \'a zs Sz \

    R \'o bert Busa-Fekete, Eyke H \"u llermeier, and Bal \'a zs Sz \"o r \'e nyi. Preference-based R ank E licitation using S tatistical M odels: T he C ase of M allows. In Proceedings of The 31st International Conference on Machine Learning, volume 32, 2014

    work page 2014

  5. [5]

    Appell F unctions and M ultiple A verages

    BC Carlson. Appell F unctions and M ultiple A verages. SIAM Journal on Mathematical Analysis, 2 0 (3): 0 420--430, 1971

    work page 1971

  6. [6]

    A S equential P article F ilter M ethod for S tatic M odels

    Nicolas Chopin. A S equential P article F ilter M ethod for S tatic M odels. Biometrika, 89 0 (3): 0 539--552, 2002

    work page 2002

  7. [7]

    Click M odels for W eb S earch

    Aleksandr Chuklin, Ilya Markov, and Maarten de Rijke. Click M odels for W eb S earch. Synthesis Lectures on Information Concepts, Retrieval, and Services, 7 0 (3): 0 1--115, 2015

    work page 2015

  8. [8]

    Deep N eural N etworks for Y ou T ube R ecommendations

    Paul Covington, Jay Adams, and Emre Sargin. Deep N eural N etworks for Y ou T ube R ecommendations. In Proceedings of the 10th ACM Conference on Recommender Systems, New York, NY, USA, 2016

    work page 2016

Show all 45 references
  1. [9]

    An E xperimental C omparison of C lick P osition- B ias M odels

    Nick Craswell, Onno Zoeter, Michael Taylor, and Bill Ramsey. An E xperimental C omparison of C lick P osition- B ias M odels. In Proceedings of the 2008 I nternational C onference on W eb S earch and D ata M ining , pages 87--94. ACM, 2008

    2008

  2. [10]

    A B ayesian A pproach to P aired C omparison E xperimentation

    Roger R Davidson and Daniel L Solomon. A B ayesian A pproach to P aired C omparison E xperimentation. Biometrika, 60 0 (3): 0 477--487, 1973

    1973

  3. [11]

    Multiple H ypergeometric F unctions: P robabilistic I nterpretations and S tatistical U ses

    James M Dickey. Multiple H ypergeometric F unctions: P robabilistic I nterpretations and S tatistical U ses. Journal of the American Statistical Association, 78 0 (383): 0 628--637, 1983

    1983

  4. [12]

    Bayesian M ethods for C ensored C ategorical D ata

    James M Dickey, Jhy-Ming Jiang, and Joseph B Kadane. Bayesian M ethods for C ensored C ategorical D ata. Journal of the American Statistical Association, 82 0 (399): 0 773--781, 1987

    1987

  5. [13]

    Maximum S election and R anking under N oisy C omparisons

    Moein Falahatgar, Alon Orlitsky, Venkatadheeraj Pichapati, and Ananda Theertha Suresh. Maximum S election and R anking under N oisy C omparisons. In International Conference on Machine Learning, pages 1088--1096, 2017

    2017

  6. [14]

    Sampling-based A pproaches to C alculating M arginal D ensities

    Alan E Gelfand and Adrian FM Smith. Sampling-based A pproaches to C alculating M arginal D ensities. Journal of the American Statistical Association, 85 0 (410): 0 398--409, 1990

    1990

  7. [15]

    Adaptive R ejection M etropolis S ampling within G ibbs S ampling

    Wally R Gilks, NG Best, and KKC Tan. Adaptive R ejection M etropolis S ampling within G ibbs S ampling. Journal of the Royal Statistical Society: Series C (Applied Statistics), 44 0 (4): 0 455--472, 1995

    1995

  8. [16]

    Thompson S ampling for C omplex O nline P roblems

    Aditya Gopalan, Shie Mannor, and Yishay Mansour. Thompson S ampling for C omplex O nline P roblems. In Proceedings of the 31st International Conference on International Conference on Machine Learning, pages 100--108, 2014

    2014

  9. [17]

    Bayesian I nference for P lackett- L uce R anking M odels

    John Guiver and Edward Snelson. Bayesian I nference for P lackett- L uce R anking M odels. In P roceedings of the 26th A nnual I nternational C onference on M achine learning , pages 377--384. ACM, 2009

    2009

  10. [18]

    A G eneralization of the D irichlet D istribution

    Robin KS Hankin. A G eneralization of the D irichlet D istribution. Journal of Statistical Software, 33 0 (11): 0 1--18, 2010

    2010

  11. [19]

    Active R anking U sing P airwise C omparisons

    Kevin G Jamieson and Robert Nowak. Active R anking U sing P airwise C omparisons. In Advances in Neural Information Processing Systems, pages 2240--2248, 2011

    2011

  12. [20]

    Degenerate F eedback L oops in R ecommender S ystems

    Ray Jiang, Silvia Chiappa, Tor Lattimore, Andras Agyorgy, and Pushmeet Kohli. Degenerate F eedback L oops in R ecommender S ystems. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, 2019

    2019

  13. [21]

    Computation of C arlson's M ultiple H ypergeometric F unction R for B ayesian A pplications

    Thomas J Jiang, Joseph B Kadane, and James M Dickey. Computation of C arlson's M ultiple H ypergeometric F unction R for B ayesian A pplications. Journal of Computational and Graphical Statistics, 1 0 (3): 0 231--251, 1992

    1992

  14. [22]

    Regret L ower B ound and O ptimal A lgorithm in D ueling B andit P roblem

    Junpei Komiyama, Junya Honda, Hisashi Kashima, and Hiroshi Nakagawa. Regret L ower B ound and O ptimal A lgorithm in D ueling B andit P roblem. In Conference on Learning Theory, pages 1141--1154, 2015

    2015

  15. [23]

    Bandit A lgorithms

    Tor Lattimore and Csaba Szepesv \'a ri. Bandit A lgorithms . Cambridge (to appear), 2019

    2019

  16. [24]

    Toprank: A P ractical A lgorithm for O nline S tochastic R anking

    Tor Lattimore, Branislav Kveton, Shuai Li, and Csaba Szepesvari. Toprank: A P ractical A lgorithm for O nline S tochastic R anking. In Advances in Neural Information Processing Systems, pages 3949--3958, 2018

    2018

  17. [25]

    An A lternative B ayesian A pproach to the B radley- T erry M odel for P aired C omparisons

    Tom Leonard. An A lternative B ayesian A pproach to the B radley- T erry M odel for P aired C omparisons. Biometrics, pages 121--132, 1977

    1977

  18. [26]

    Causal I nference for R ecommendation

    Dawen Liang, Laurent Charlin, and David M Blei. Causal I nference for R ecommendation. In UAI Workshop on Causation: Foundation to Application, 2016

    2016

  19. [27]

    Variational A utoencoders for C ollaborative F iltering

    Dawen Liang, Rahul G Krishnan, Matthew D Hoffman, and Tony Jebara. Variational A utoencoders for C ollaborative F iltering. In Proceedings of the 2018 World Wide Web Conference on World Wide Web, pages 689--698. International World Wide Web Conferences Steering Committee, 2018

    2018

  20. [28]

    Learning to R ank for I nformation R etrieval

    Tie-Yan Liu. Learning to R ank for I nformation R etrieval. Foundations and Trends in Information Retrieval , 3 0 (3): 0 225--331, 2009

    2009

  21. [29]

    Ducan Luce

    R. Ducan Luce. Individual C hoice B ehavior . Wiley, 1959

    1959

  22. [30]

    Active L earning for T op- k R ank A ggregation from N oisy C omparisons

    Soheil Mohajer, Changho Suh, and Adel Elmahdy. Active L earning for T op- k R ank A ggregation from N oisy C omparisons. In International Conference on Machine Learning, pages 2488--2497, 2017

    2017

  23. [31]

    Radford M. Neal. Slice S ampling. The Annals of Statistics, 31 0 (3): 0 705--767, 2003

    2003

  24. [32]

    The F ilter B ubble: W hat the I nternet is hiding from you

    Eli Pariser. The F ilter B ubble: W hat the I nternet is hiding from you . Penguin UK, 2011

    2011

  25. [33]

    The A nalysis of P ermutations

    Robin L Plackett. The A nalysis of P ermutations. Applied Statistics, pages 193--202, 1975

    1975

  26. [34]

    Transitivity of P references

    Michel Regenwetter, Jason Dana, and Clintin P Davis-Stober. Transitivity of P references. Psychological Review, 118 0 (1): 0 42, 2011

    2011

  27. [35]

    Battle of B andits

    Aadirupa Saha and Aditya Gopalan. Battle of B andits. In Proceedings of the Thirty-Forth Conference on Uncertainty in Artificial Intelligence, UAI, pages 06--10, 2018

    2018

  28. [36]

    Human I nteraction with R ecommendation S ystems

    Sven Schmit and Carlos Riquelme. Human I nteraction with R ecommendation S ystems. In Proceedings of the 21th International Conference on Artificial Intelligence and Statistics, 2018

    2018

  29. [37]

    Intransitive C ycles: R ational C hoice or R andom E rror? an A nswer based on E stimation of E rror R ates with E xperimental D ata

    Barry Sopher and Gary Gigliotti. Intransitive C ycles: R ational C hoice or R andom E rror? an A nswer based on E stimation of E rror R ates with E xperimental D ata. Theory and decision, 35 0 (3): 0 311--336, 1993

    1993

  30. [38]

    Multi-dueling B andits with D ependent A rms

    Yanan Sui, Vincent Zhuang, Joel W Burdick, and Yisong Yue. Multi-dueling B andits with D ependent A rms. In Proceedings of the Thirty-Forth Conference on Uncertainty in Artificial Intelligence, UAI, 2017

    2017

  31. [39]

    Advancements in D ueling B andits

    Yanan Sui, Masrour Zoghi, Katja Hofmann, and Yisong Yue. Advancements in D ueling B andits. In IJCAI, pages 5502--5510, 2018

    2018

  32. [40]

    o r \'e nyi, R \'o bert Busa-Fekete, Adil Paul, and Eyke H \

    Bal \'a zs Sz \"o r \'e nyi, R \'o bert Busa-Fekete, Adil Paul, and Eyke H \"u llermeier. Online R ank elicitation for P lackett-luce: A D ueling B andits A pproach. In Advances in Neural Information Processing Systems, pages 604--612, 2015

    2015

  33. [41]

    On the L ikelihood that O ne U nknown P robability E xceeds A nother in V iew of the E vidence of T wo S amples

    William R Thompson. On the L ikelihood that O ne U nknown P robability E xceeds A nother in V iew of the E vidence of T wo S amples. Biometrika, 25 0 (3/4): 0 285--294, 1933

    1933

  34. [42]

    Double T hompson S ampling for D ueling B andits

    Huasen Wu and Xin Liu. Double T hompson S ampling for D ueling B andits. In Advances in Neural Information Processing Systems, pages 649--657, 2016

    2016

  35. [43]

    Collaborative C ompetitive F iltering: L earning R ecommender using C ontext of U ser C hoice

    Shuang-Hong Yang, Bo Long, Alexander J Smola, Hongyuan Zha, and Zhaohui Zheng. Collaborative C ompetitive F iltering: L earning R ecommender using C ontext of U ser C hoice. In Proceedings of the 34th international ACM SIGIR conference on Research and development in Informatio...

    2011

  36. [44]

    The R elationship between luce's C hoice A xiom, T hurstone's T heory of C omparative J udgment, and the D ouble E xponential D istribution

    John I Yellott Jr. The R elationship between luce's C hoice A xiom, T hurstone's T heory of C omparative J udgment, and the D ouble E xponential D istribution. Journal of Mathematical Psychology, 15 0 (2): 0 109--144, 1977

    1977

  37. [45]

    The k-armed D ueling B andits P roblem

    Yisong Yue, Josef Broder, Robert Kleinberg, and Thorsten Joachims. The k-armed D ueling B andits P roblem. Journal of Computer and System Sciences, 78 0 (5): 0 1538--1556, 2012

    2012

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.