Pith. sign in

REVIEW 3 major objections 5 minor 47 references

A Non-Parametric Choice Model That Learns How Users Choose Between Recommended Options

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 LCM4Rec learns the distribution of random errors in users' choices from data, so a recommender can infer both what users prefer and how they choose without fixing a choice model in advance.

desk verdict Useful, reproducible methods paper whose 'recovers the choice model' claim is undercut by an unaddressed scale invariance; predictive results still stand. read the letter →

arxiv 2507.20035 v1 pith:ACI4VYRB submitted 2025-07-26 cs.IR

Thorsten Krause , Harrie Oosterhuis This is my paper · ORCID
classification cs.IR
keywords RecommendersystemsUserchoicemodelingExposurebiasNon-parametricmodelsKerneldensityestimationMultinomiallogitExponomialmodelprobability
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

This paper proposes LCM4Rec, a non-parametric choice model for recommender systems that learns the distribution of random errors in users' choices instead of assuming one upfront, such as the Gumbel noise behind the multinomial logit model. The authors' central claim is that, by approximating the error distribution with a differentiable kernel density estimator, LCM4Rec can simultaneously recover what users prefer and how they choose among the options shown, making it applicable to any choice model whose error terms are independent and identically distributed. They show on synthetic data that it recovers the shape of the true error distribution, remains competitive across different generating models, and is less affected by exposure bias coming from competition than standard parametric models. If the claim holds, recommender systems no longer need to guess which choice model users follow, and preference inference becomes more reliable under mismatch.

What carries the argument

The load-bearing object is a kernel-density estimate of the error distribution built from sigmoid kernels, $\hat{F}_K(x)=\sum_{k=1}^K w_k\sigma((x-x_k)/h_k)$, a weighted average of translated and scaled sigmoid functions whose derivative gives a smooth, differentiable pdf. This family is dense enough to approximate any continuous, strictly monotone cdf with bounded support, yet smooth enough that both its parameters and the utility estimates can be optimized by gradient descent. The likelihood is then approximated by drawing samples from each sigmoid kernel through its inverse cdf and averaging, with a third-order Taylor correction for the bias introduced by taking the logarithm of the Monte Carlo estimate.

What would settle it

Train LCM4Rec on data generated under a model its assumptions exclude, such as nested logit with correlated alternatives, or with choice sets where one alternative is unobserved. If the recovered choice probabilities and preference rankings do not converge to the true ones as the number of observations and kernels grows, then the IID-error and fully-observed-choice-set assumptions are load-bearing rather than cosmetic.

Watch

Extended reading notes

Core claim

LCM4Rec's core discovery is that the choice probability integral of Equation (3) can be optimized without parametric assumptions by replacing the unknown error cdf $F_\epsilon$ with a weighted sum of sigmoid kernels, $\hat{F}_K(x)=\sum_{k=1}^K w_k\sigma((x-x_k)/h_k)$, and by approximating the likelihood gradient through Monte Carlo samples drawn from those kernels. Because this function family is dense enough to approximate any continuous, strictly monotone error cdf with bounded support, the model can in principle approximate any choice model with IID errors. The paper reports that on synthetic data generated under Gumbel, signed-exponential, and Gaussian-mixture errors, LCM4Rec is the only method with consistently competitive choice probabilities and accurately recovers the error distribution's shape; it is also the least affected by exposure bias from competition, even when the parametric models' error assumptions match the data-generating process.

Load-bearing premise

The load-bearing premise is that every user-item error term is drawn independently from one unknown distribution and that the full set of options a user saw is recorded in the data; if real choices involve correlated alternatives or unobserved options, the learned error distribution and the exposure-bias robustness may not carry over.

Editorial extensions

If this is right

  • The method removes the need to choose or benchmark a parametric choice model in advance, because the error distribution is estimated from the same logged interactions used to learn preferences.
  • Users' choice behavior can be studied directly: the estimated cdf and pdf can reveal whether choices are conservative, non-conservative, or multimodal, as the Gaussian-mixture experiment demonstrates.
  • Exposure bias from competition, which parametric models suffer when their error assumption is wrong, is reduced because the estimated error distribution adapts to the data.
  • The architecture is differentiable and therefore compatible with gradient-descent training of recommender embeddings, unlike earlier non-parametric estimators that relied on sorting, linear programming, or discrete optimization.

Reading between the lines

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

  • Because the current model is restricted to IID errors, extending it to nested or cross-nested structures would require modeling correlations between alternatives; the paper flags this as future work, but the same kernel machinery could be adapted with a factorized or copula-based error model.
  • The exposure-bias results suggest a testable hypothesis: learning the error density and preferences jointly may implicitly regularize the model against competition-induced confounding, and comparing LCM4Rec with a variant that uses a fixed but correctly specified error distribution would isolate whether the benefit comes from learning the density or from the non-parametric family itself.
  • On partially observed real-world choice sets, the estimated error distribution may absorb the effect of missing alternatives; measuring how inferred preferences shift when a no-choice option is added to the choice set would indicate how much of the method's performance depends on full observability.
  • The reported $O(K^2 N S |C|)$ complexity with small $K$ is practical, but larger choice sets or online inference would likely require alternative sampling or sub-sampling strategies, which the paper does not address.
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 LCM4Rec, a differentiable non-parametric discrete choice model for recommender systems. It models the (assumed IID) error distribution of a random-utility model as a mixture of sigmoid kernels whose weights, locations, and widths are learned, and it jointly learns user/item embeddings and item constants by maximizing a Monte Carlo approximation of the log-likelihood (Eqs. 5–17). The authors prove an approximation theorem for the kernel family, describe gradient-based optimization with a bias correction, and report synthetic experiments comparing LCM4Rec with multinomial logit, exponomial, and univariate baselines under three error distributions. The experiments address three questions: accuracy/robustness of choice probabilities, recovery of the true error distribution, and robustness to exposure bias. The paper claims that LCM4Rec simultaneously infers preferences and the choice model, is robust across true choice models, and resists exposure bias better than parametric models.

Significance. If correct, the contribution is significant for recommendation choice modeling: it would remove the need to commit a priori to a parametric error distribution, and the synthetic evidence suggests that choice-probability prediction and exposure-bias robustness can be improved by learning the error distribution. The paper's strengths include a clear research-question structure, repeated synthetic trials (100 repetitions), public code/data, and a model that is straightforward to optimize. The central 'recovery of the choice model' claim, however, is currently not supported because of a scale-invariance of the likelihood (detailed below); with a proper identification argument or a weakened claim, the manuscript would be a useful step toward data-driven choice modeling.

major comments (3)
  1. [§3.4, Eq. (8)–(9); Table 3 and Figure 1] The likelihood is invariant to a common scaling of utilities and errors, and the c_j-clipping in Section 3.4 does not remove this invariance. For any fitted parameter vector and any a>1 with l/a within the allowed range and c_j/a in [0,1], the transformation u_i'=u_i/sqrt(a), v_j'=v_j/sqrt(a), c_j'=c_j/a, l'=l/a, h_k'=h_k/a (with unchanged w_k, and x_k scaled to x_k/a) yields V'_ij=V_ij/a and an error distribution F'(x)=F(a x), hence U'_ij=U_ij/a. Every ranking and every choice probability is unchanged, so the negative log-likelihood is identical and the model family and constraints are satisfied. Consequently 'the most likely choice model' is defined only up to scale (and location), and RQ2's evaluation is not valid as reported: the estimates in Table 3 and Figure 1 are aligned only for translation, not scale, so the low KLD values do not demonstrate recovery of the true error distribution. The authors should impose an explicit scale normalization (e.g., fixing the variance or a quantile of F_hat), prove that the normalized model is identified, and report both the choice-probability and the normalized error-recovery results.
  2. [§3.1, Theorem 1] The universal-approximation claim is not established by the proof as written. The theorem is stated only for continuous strictly monotone cdfs with bounded support, yet the experiments use Gumbel, signed exponential, and Gaussian mixture errors, all with unbounded support; the text after the theorem repeats the bounded-support restriction, which is weaker than the paper's claim of applicability to 'any choice model' with IID errors. Within the proof, the error term xi in Eq. (12) is asserted to vanish as h→0 without a quantitative bound, the instruction to set 'the width lambda large enough' is unclear because lambda controls l, not kernel width, and the final inequality uses F(x_{k*+1}) even when k*=K. The proof should be rewritten with explicit quantifiers, a bound on xi, and a statement covering the distributions actually used, or the claim should be restricted accordingly.
  3. [§3.2; §5.2] The paper provides no consistency or identifiability result for the sieve maximum-likelihood estimator. Section 3.2 asserts that minimizing the approximated NLL 'can potentially recover any underlying choice model,' but there is no argument that the global maximizer of the population likelihood isolates the true (V,F) as N grows, nor any guidance on how K should grow with N. This gap is closely tied to the scale invariance in the first comment; even after fixing scale, a consistency theorem or a precise statement of what is recoverable (e.g., choice probabilities and the error distribution up to scale/location) is needed to support RQ2 and the abstract's 'simultaneously infers what users prefer and how they make choices.'
minor comments (5)
  1. [§1] The sentence 'One prominent class of models than accounts for dependencies' should read 'that accounts'; similarly, 'both learns users preferences as how users choose' is ungrammatical and should be rephrased.
  2. [§3.1, Eq. (12)] Equation (12) has typesetting issues: 'F(x_k-1)' should be 'F(x_{k-1})', and the underbrace annotation is incomplete, making the displayed equality hard to read.
  3. [Table 2] The significance notation is unclear: stars appear next to individual values, but the footnote does not specify which pairwise comparisons are tested or how the 'next worse model' is defined; please clarify with a table note.
  4. [Figure 1 caption] The caption says estimates are shifted to minimize KLD to the true distribution, but it does not state how the KLD is computed over the support or whether any scale alignment is performed; please add these details.
  5. [§4.2] The phrase 'resulting in 1.9 times as much exposure per item' is not defined; please specify how this factor is computed from the described exposure process.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: recovery is tested against held-out choices and known generative parameters, and self-citations to prior work are not load-bearing.

full rationale

The paper's derivation chain is self-contained. Choice probabilities are defined from an IID error-term random utility model (Eq. 3); the error CDF is approximated by a differentiable sigmoid-kernel family (Eqs. 5-8); and model parameters, including the error distribution, are fit by minimizing a Monte Carlo approximation of the negative log-likelihood (Eqs. 16-17). The target quantities—the true error distribution and held-out choice probabilities—are not fed into the loss; they are only used after fitting to measure recovery accuracy (Table 3, Figure 1) and predictive performance (Table 2). Thus, the 'recovery' claim is an empirical identification claim evaluated against generative ground truth, not a fitted input renamed as a prediction. The only self-citations are to the authors' prior exposure-bias framework [29], used for defining exposure bias and constructing experimental protocols; the paper explicitly notes that the prior proof assumed a correctly specified choice model and does not rely on it to establish LCM4Rec's robustness, which is instead evaluated empirically in Experiment 2. The scale-invariance/identification issue discussed in Section 3.4 is a legitimate identifiability concern, but it is not a circular reduction: the likelihood is not definitionally equal to its inputs, and the paper's invariant predictions (choice probabilities, rankings) are still evaluated on held-out data. No circular step satisfying the quoted-reduction standard was found.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the IID error assumption, known choice sets, and a matrix-factorization utility form. No new physical or mathematical entities are introduced. The kernel count K, Monte Carlo sample size S, and embedding dimension m are hand-chosen hyperparameters rather than fitted scientific constants.

free parameters (3)
  • Kernel count K = 5
    Number of sigmoid kernels in the density estimator; chosen by hand after finding no benefit beyond K=5; controls approximation quality of the error distribution.
  • Monte Carlo samples S = 5
    Samples per kernel per observation for gradient approximation; chosen by hand; affects variance of the NLL estimator.
  • Embedding dimension m = 3
    Set to match the synthetic generation; in practice a hyperparameter of the utility model V_ij = u_i dot v_j + c_j.
assumptions (5)
  • domain assumption Error terms epsilon_ij are IID with unknown distribution F
    Used in Equation 3 and throughout Section 3.1; excludes nested logit and other correlated-error models, acknowledged by the authors.
  • domain assumption Each user chooses the item with highest utility U_ij = V_ij + epsilon_ij
    Foundation of discrete choice models, stated in Section 2.1.
  • domain assumption The full choice set C_n for each observation is known
    Stated as a limitation in Section 4.3; real-world choice sets can be partially unobserved.
  • domain assumption Deterministic utility has the form V_ij = u_i dot v_j + c_j
    Equation 9 in Section 3.1; a matrix factorization assumption standard in recommender systems.
  • standard math The sigmoid-mixture family can approximate any continuous, strictly monotone cdf with finite support
    Theorem 1 in Section 3.1; the proof is a sketch with an unquantified error term, and the finite-support assumption does not cover the unbounded Gumbel and exponential distributions used in experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Non-Parametric Choice Model That Learns How Users Choose Between Recommended Options." pith.science (2026). https://pith.science/paper/ACI4VYRB

@misc{pith2026250720035,
  author       = {Pith},
  title        = {Pith review of: A Non-Parametric Choice Model That Learns How Users Choose Between Recommended Options},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ACI4VYRB}},
  note         = {Machine review of arXiv:2507.20035}
}
read the original abstract

Choice models predict which items users choose from presented options. In recommendation settings, they can infer user preferences while countering exposure bias. In contrast with traditional univariate recommendation models, choice models consider which competitors appeared with the chosen item. This ability allows them to distinguish whether a user chose an item due to preference, i.e., they liked it; or competition, i.e., it was the best available option. Each choice model assumes specific user behavior, e.g., the multinomial logit model. However, it is currently unclear how accurately these assumptions capture actual user behavior, how wrong assumptions impact inference, and whether better models exist. In this work, we propose the learned choice model for recommendation (LCM4Rec), a non-parametric method for estimating the choice model. By applying kernel density estimation, LCM4Rec infers the most likely error distribution that describes the effect of inter-item cannibalization and thereby characterizes the users' choice model. Thus, it simultaneously infers what users prefer and how they make choices. Our experimental results indicate that our method (i) can accurately recover the choice model underlying a dataset; (ii) provides robust user preference inference, in contrast with existing choice models that are only effective when their assumptions match user behavior; and (iii) is more resistant against exposure bias than existing choice models. Thereby, we show that learning choice models, instead of assuming them, can produce more robust predictions. We believe this work provides an important step towards better understanding users' choice behavior.

Figures

Figures reproduced from arXiv: 2507.20035 by the authors.

Figure 1
Figure 1. Estimated versus true cdfs and pdfs and KLD score between the mean estimate and the true distribution, smoothed to [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Exposure bias w.r.t. items’ mean true ranks. Regression curves and CIs based on LOWESS [ [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 28 canonical work pages

  1. [1]

    Himan Abdollahpouri and Masoud Mansoury. 2020. Multi-sided Exposure Bias in Recommendation. arXiv:2006.15772 [cs.IR] https://arxiv.org/abs/2006.15772

    arXiv 2020

  2. [2]

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

  3. [3]

    Aydın Alptekinoğlu and John H Semple. 2016. The exponomial choice model: A new alternative for assortment and price optimization. Operations Research 64, 1 (2016), 79–93

    work page 2016

  4. [4]

    Moshe Ben-Akiva and Michel Bierlaire. 1999. Discrete Choice Methods and their Applications to Short Term Travel Decisions . Springer US, Boston, MA, 5–33. doi:10.1007/978-1-4615-5203-1_2

    work page doi:10.1007/978-1-4615-5203-1_2 1999

  5. [5]

    Moshe E Ben-Akiva and Steven R Lerman. 1985. Discrete choice analysis: theory and application to travel demand . Vol. 9. MIT press, Boston, USA

    work page 1985

  6. [6]

    Benson, Ravi Kumar, and Andrew Tomkins

    Austin R. Benson, Ravi Kumar, and Andrew Tomkins. 2016. On the Relevance of Irrelevant Alternatives. In Proceedings of the 25th International Conference on World Wide Web (Montréal, Québec, Canada) (WWW ’16). International World Wide Web Conferences Steering Committee, Republic and Canton of Geneva, CHE, 963–973. doi:10.1145/2872427.2883025 9 RecSys ’25, ...

    arXiv 2016

  7. [7]

    Robert C Blattberg and Kenneth J Wisniewski. 1989. Price-induced patterns of competition. Marketing science 8, 4 (1989), 291–309

    work page 1989

  8. [8]

    Richard A Briesch, Pradeep K Chintagunta, and Rosa L Matzkin. 2010. Nonpara- metric discrete choice models with unobserved heterogeneity.Journal of Business & Economic Statistics 28, 2 (2010), 291–307

    work page 2010

Show all 47 references
  1. [9]

    Sebastian Büscher and Dietmar Bauer. 2024. Weighting strategies for pairwise composite marginal likelihood estimation in case of unbalanced panels and unaccounted autoregressive structure of the errors. Transportation research part B: methodological 181 (2024), 102890

    2024

  2. [10]

    Busemeyer and Jörg Rieskamp

    Jerome R. Busemeyer and Jörg Rieskamp. 2014. Chapter 3: Psychological research and theories on preferential choice. In Handbook of Choice Modelling . Edward Elgar Publishing, Cheltenham, UK. doi:10.4337/9781781003152.00008

    2014

  3. [11]

    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 Systems 41, 3 (2023), 1–39

    2023

  4. [12]

    Khai Chiong, Yu-Wei Hsieh, and Matthew Shum. 2017. Counterfactual Evaluation in Semiparametric Multinomial Choice Models

    2017

  5. [13]

    William S Cleveland. 1979. Robust locally weighted regression and smoothing scatterplots. Journal of the American statistical association74, 368 (1979), 829–836

    1979

  6. [14]

    James L Corner and Craig W Kirkwood. 1991. Decision analysis applications in the operations research literature, 1970–1989. Operations Research 39, 2 (1991), 206–219

    1991

  7. [15]

    Cosslett

    Stephen R. Cosslett. 1983. Distribution-Free Maximum Likelihood Estimator of the Binary Choice Model. Econometrica 51, 3 (1983), 765–782. http://www.jstor. org/stable/1912157

    1983

  8. [16]

    Abhisek Dash, Abhijnan Chakraborty, Saptarshi Ghosh, Animesh Mukherjee, and Krishna P. Gummadi. 2021. When the Umpire is also a Player: Bias in Private Label Product Recommendations on E-commerce Marketplaces. In Proceedings of the 2021 ACM Conference on Fairness, Accountabili...

    2021

  9. [17]

    James Durbin and Siem Jan Koopman. 1997. Monte Carlo maximum likelihood estimation for non-Gaussian state space models.Biometrika 84, 3 (1997), 669–684

    1997

  10. [18]

    Simen Eide, David S Leslie, Arnoldo Frigessi, Joakim Rishaug, Helge Jenssen, and Sofie Verrewaere. 2021. FINN.no Slates Dataset: A new Sequential Dataset Logging Interactions, all Viewed Items and Click Responses/No-Click for Recom- mender Systems Research. In Proceedings of t...

    2021

  11. [19]

    Andres Ferraro, Dmitry Bogdanov, Xavier Serra, and Jason Yoon. 2019. Artist and style exposure bias in collaborative filtering based music recommendations. arXiv:1911.04827 [cs.IR] https://arxiv.org/abs/1911.04827

    2019 arXiv

  12. [20]

    Duran, Pere Gilabert, Santi Seguí, and Jordi Vitrià

    Paula G. Duran, Pere Gilabert, Santi Seguí, and Jordi Vitrià. 2024. Overcoming diverse undesired effects in recommender systems: A deontological approach. ACM Transactions on Intelligent Systems and Technology 15, 4 (2024), 1–23

    2024

  13. [21]

    Milad Haghani, Michiel CJ Bliemer, and David A Hensher. 2021. The landscape of econometric discrete choice modelling research. Journal of choice modelling 40 (2021), 100303

    2021

  14. [22]

    Keisuke Hirano. 2002. Semiparametric Bayesian inference in autoregressive panel data models. Econometrica 70, 2 (2002), 781–799

    2002

  15. [23]

    Horowitz

    Joel L. Horowitz. 1992. A Smoothed Maximum Score Estimator for the Binary Response Model. Econometrica 60, 3 (1992), 505–531. http://www.jstor.org/ stable/2951582

    1992

  16. [24]

    Yifan Hu, Yehuda Koren, and Chris Volinsky. 2008. Collaborative Filtering for Implicit Feedback Datasets. In 2008 Eighth IEEE International Conference on Data Mining. IEEE, Piscataway, NJ, 263–272. doi:10.1109/ICDM.2008.22

    2008 doi

  17. [25]

    Hidehiko Ichimura and T Scott Thompson. 1998. Maximum likelihood estimation of a binary choice model with random coefficients of unknown distribution. Journal of Econometrics 86, 2 (1998), 269–295

    1998

  18. [26]

    Ray Jiang, Silvia Chiappa, Tor Lattimore, András György, and Pushmeet Kohli

  19. [27]

    Ho Yin Kan, Dennis Wong, and Keith Chau. 2024. A Personalized Flight Recom- mender System Based on Link Prediction in Aviation Data. IEEE Access 12 (2024), 36961–36973. doi:10.1109/ACCESS.2024.3369487

    2024

  20. [28]

    Klein and Richard H

    Roger W. Klein and Richard H. Spady. 1993. An Efficient Semiparametric Estimator for Binary Response Models. Econometrica 61, 2 (1993), 387–421. http://www.jstor.org/stable/2951556

    1993

  21. [29]

    Beinke, Gerrit Y

    Thorsten Krause, Alina Deriyeva, Jan H. Beinke, Gerrit Y. Bartels, and Oliver Thomas. 2024. Mitigating Exposure Bias in Recommender Systems—A Compara- tive Analysis of Discrete Choice Models. ACM Trans. Recomm. Syst. 3, 2, Article 19 (Nov. 2024), 37 pages. doi:10.1145/3641291

    2024 doi

  22. [30]

    Jiayu Li, Aixin Sun, Weizhi Ma, Peijie Sun, and Min Zhang. 2024. Right Tool, Right Job: Recommendation for Repeat and Exploration Consumption in Food Delivery. In Proceedings of the 18th ACM Conference on Recommender Systems (Bari, Italy) (RecSys ’24). Association for Computin...

    2024

  23. [31]

    Stan Lipovetsky and Michael Conklin. 2014. Finding items cannibalization and synergy by BWS data. Journal of choice modelling 12 (2014), 1–9

    2014

  24. [32]

    Charles F Manski. 1975. Maximum score estimation of the stochastic utility model of choice. Journal of econometrics 3, 3 (1975), 205–228

    1975

  25. [33]

    Masoud Mansoury, Himan Abdollahpouri, Mykola Pechenizkiy, Bamshad Mobasher, and Robin Burke. 2020. Feedback Loop and Bias Amplification in Recommender Systems. In Proceedings of the 29th ACM International Confer- ence on Information & Knowledge Management (Virtual Event, Irela...

    2020

  26. [34]

    Rosa L. Matzkin. 1992. Nonparametric and Distribution-Free Estimation of the Binary Threshold Crossing and The Binary Choice Models. Econometrica 60, 2 (1992), 239–270. http://www.jstor.org/stable/2951596

    1992

  27. [35]

    Aleksandr Vladimirovich Petrov and Craig Macdonald. 2023. gSASRec: Reducing Overconfidence in Sequential Recommendation Trained with Negative Sampling. In Proceedings of the 17th ACM Conference on Recommender Systems (Singapore, Singapore) (RecSys ’23). Association for Computi...

    2023

  28. [36]

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme

  29. [37]

    Harald Steck, Linas Baltrunas, Ehtsham Elahi, Dawen Liang, Yves Raimond, and Justin Basilico. 2021. Deep learning for recommender systems: A Netflix case study. AI Magazine 42, 3 (2021), 7–18

    2021

  30. [38]

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang

  31. [39]

    Pietro Tebaldi, Alexander Torgovitsky, and Hanbin Yang. 2023. Nonparametric estimates of demand in the california health insurance exchange. Econometrica 91, 1 (2023), 107–146

    2023

  32. [40]

    Kenneth E. Train. 2009.Discrete Choice Methods with Simulation(2nd ed.). Number 9780521766555 in Cambridge Books. Cambridge University Press, Cambridge, UK

    2009

  33. [41]

    Qi Wan, Xiangnan He, Xiang Wang, Jiancan Wu, Wei Guo, and Ruiming Tang

  34. [42]

    In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (Beijing, China) (CIKM ’19)

    BERT4Rec: Sequential Recommendation with Bidirectional Encoder Rep- resentations from Transformer. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (Beijing, China) (CIKM ’19). Association for Computing Machinery, New York, NY, US...

  35. [47]

    Smola, Hongyuan Zha, and Zhaohui Zheng

    Shuang-Hong Yang, Bo Long, Alexander J. Smola, Hongyuan Zha, and Zhaohui Zheng. 2011. Collaborative competitive filtering: learning recommender using context of user choice. In Proceedings of the 34th International ACM SIGIR Con- ference on Research and Development in Informat...

    2011

  36. [2009]

    In Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence (Montreal, Quebec, Canada) (UAI ’09)

    BPR: Bayesian personalized ranking from implicit feedback. In Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence (Montreal, Quebec, Canada) (UAI ’09). AUAI Press, Arlington, Virginia, USA, 452–461

  37. [2019]

    In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society (Honolulu, HI, USA) (AIES ’19)

    Degenerate Feedback Loops in Recommender Systems. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society (Honolulu, HI, USA) (AIES ’19). Association for Computing Machinery, New York, NY, USA, 383–390. doi:10.1145/3306618.3314288

    2019

  38. [2020]

    In Proceedings of the 14th ACM Conference on Recommender Systems (Virtual Event, Brazil)(RecSys ’20)

    The Connection Between Popularity Bias, Calibration, and Fairness in Recommendation. In Proceedings of the 14th ACM Conference on Recommender Systems (Virtual Event, Brazil)(RecSys ’20). Association for Computing Machinery, New York, NY, USA, 726–731. doi:10.1145/3383313.3418487

  39. [2022]

    InProceedings of the ACM Web Conference 2022 (Virtual Event, Lyon, France) (WWW ’22)

    Cross Pairwise Ranking for Unbiased Item Recommendation. InProceedings of the ACM Web Conference 2022 (Virtual Event, Lyon, France) (WWW ’22) . Association for Computing Machinery, New York, NY, USA, 2370–2378. doi:10. 1145/3485447.3512010

    2022

Pith tools

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