Pith. sign in

REVIEW 2 major objections 5 minor 39 references

Recommendation and Temptation

T0 review · 2 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read When a platform knows everything except which outside option a user will face, the optimal strategy is locally greedy: each round, recommend the available item with the highest expected enrichment against that outside option.

desk verdict Clean behavioral model and a plausible greedy theorem, but the theorem needs a stationarity assumption and the empirical claims are in-simulation only. read the letter →

arxiv 2412.10595 v2 pith:BI4ZR3S6 submitted 2024-12-13 cs.IR cs.CYcs.GT

Md Sanzeed Anwar , Paramveer S. Dhillon , Grant Schoenebeck This is my paper · ORCID
classification cs.IRcs.CYcs.GT
keywords RecommenderSystemsTemptationEnrichmentOutsideOptionsConsumedGreedyStrategyUserWell-beingBehavioralModel
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 argues that recommender systems should optimize for the enrichment users actually get from what they consume, not for engagement, and that this goal is achievable despite users being pulled by temptation. It models every item as having two separate attributes for each user: enrichment, the value the user genuinely wants, and temptation, the appeal that drives the actual choice, with users choosing between recommended items and an uncertain outside option by a weighted score. The central theoretical claim is that under perfect information about items, users, and the outside-option distribution, the optimal strategy is locally greedy: each round, recommend the item whose expected enrichment against that round's outside option is largest. The paper also gives an estimation recipe that recovers enrichment and temptation from ratings plus choice data, and shows in simulations and a movie-ratings sandbox that this policy outperforms enrichment-only, temptation-only, rating-based, and click-based recommenders. If the claim holds, a simple myopic rule can steer consumption toward what users value while honestly accounting for their tendency to be tempted.

What carries the argument

The central object is the conditional enrichment $u_j(i\mid o)$, which is $u_j(i)$ when item $i$ beats outside option $o$ on choice score and $u_j(o)$ otherwise. The greedy rule takes the expectation of this quantity over the outside-option distribution and recommends the item with the largest value. The proof machinery is the reduction of any recommendation strategy to a binary tree, because the user's choice is deterministic and a set of recommendations is equivalent to its highest-choice-score item, followed by an induction on depth that swaps a strategy's first recommended item with later greedy choices and shows expected enrichment can only improve.

What would settle it

Brute-force a small finite instance, two items and two rounds, with outside options drawn i.i.d. from a fixed two-point distribution and all enrichment and temptation values known, and compare every non-greedy strategy tree against the greedy one; a single instance in which greedy's expected total enrichment is strictly smaller would refute Theorem 1.

Watch

Extended reading notes

Core claim

Formally, each user $j$ evaluates an option by the choice score $\lambda^C_j u_j(i) + (1-\lambda^C_j) v_j(i)$, where $u_j$ is enrichment, $v_j$ is temptation, and $\lambda^C_j \in [0,1]$ weights value versus appeal. In each round the user chooses the option with the highest score among the recommended items and the realized outside option. The paper's central discovery is Theorem 1: with perfect knowledge except for which outside option will be available, the expected-total-enrichment-optimal policy is locally greedy, recommending the available item $i^* = \arg\max_i \mathbb{E}_{o_j(t)}[u_j(i \mid o_j(t))]$ at every round. Here $u_j(i\mid o)$ is the enrichment received when the only alternatives are item $i$ and outside option $o$: the user consumes $i$ if its choice score is at least the outside option's score, and otherwise consumes $o$. The paper establishes this by a tree-swapping induction and then argues, via simulations and a movie-ratings sandbox, that its estimation framework lets the policy approximate this optimum from historical data.

Load-bearing premise

The theorem's proof assumes the outside-option distribution is the same in every round, so the greedy item ordering stays fixed; if outside options get systematically stronger or weaker over time, saving a tempting item for a later round could beat the greedy rule.

Editorial extensions

If this is right

  • Recommending an additional item can reduce user enrichment when the extra item is tempting but unenriching, because it crowds out a better choice.
  • A high-enrichment recommendation can fail if it is not tempting enough to beat the outside option, so consumed enrichment, not recommended enrichment, is the quantity to optimize.
  • When on-platform content is less enriching than users' outside options, the optimal policy can steer users off-platform by recommending items with low temptation.
  • The estimation framework gives a practical route to the greedy policy from historical recommendation logs plus ratings, using only broad distributional assumptions about outside options.

Reading between the lines

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

  • Beyond the paper, if the outside-option distribution changes over time, the greedy rule's fixed item ordering can break: a tempting item might be worth saving for a later round when outside options are weaker, and a lookahead policy could beat greedy.
  • Beyond the paper, the model's structure suggests a connection to bandit problems where recommendations consume items; treating the greedy rule as one-step lookahead may open extensions to exploration, multiple platforms, or learning while recommending.
  • Beyond the paper, the supply-side argument implies a testable platform-level prediction: if recommendations are chosen for consumed enrichment, content production should shift toward higher-enrichment, lower-temptation items over time, measurable in creator-side experiments.
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

2 major / 5 minor

Summary. The paper proposes a recommender-system model in which each user's choice among recommended items and an outside option is driven by a weighted combination of enrichment and temptation, while the platform's objective is to maximize expected consumed enrichment. The central theoretical claim (Theorem 1, Section 3) is that, under perfect knowledge of everything except the exact realization of the outside option, the optimal strategy is locally greedy: each round, recommend the available item maximizing the expected enrichment conditional on the item and the outside option. The paper also contributes an estimation framework (Section 4) that combines rating data and implicit choice data with distributional assumptions on outside options, and it reports simulation and MovieLens-based evaluations (Sections 5--6) claiming consistent outperformance over enrichment-based, temptation-based, ratings-based, and click-based baselines.

Significance. If the optimality theorem is correct under a clearly stated set of assumptions, the result is conceptually interesting and practically relevant: it gives a simple, myopic recommendation rule in a setting with temptation and outside options, and it sharpens the argument that engagement-based recommenders may fail to maximize user enrichment. The attempt to separate enrichment from temptation and to model off-platform alternatives is a worthwhile direction for the recommender-systems community. The paper also ships code, which is a strength. However, the theoretical claim as stated is too broad: the proof requires a stationarity assumption on the outside-option distribution that is absent from the model statement, and the empirical evaluation is largely internal to the model's own generative assumptions, so the headline claims should be read as conditional on those assumptions.

major comments (2)
  1. [Appendices] Theorem 1 is false as stated because the model does not impose stationarity of the outside-option distribution, while the proof requires it. Section 2 defines o_j(t) as 'chosen according to some probability distribution' with no statement that this distribution is the same for every t. The proof in Appendix A defines i_k as 'the kth new item the locally greedy strategy would recommend' and then performs an exchange argument over a fixed greedy ordering i_1, i_2, ...; this ordering is only well-defined if the greedy recommendation rule is time-invariant, which in turn requires the outside-option distribution to be stationary. If the distribution changes over rounds, the greedy choice in later rounds is governed by a different ranking, so the case analysis and the bijection between paths of the optimal strategy and the greedy strategy do not go through. The concern is not merely technical: for T=2, take item A with u_A=11 and choice score 100, item B with u_B=100 and choice score 50, round 1 outside score 0 with probability 0.1, 75 with probability 0.8, and 200 with probability 0.1 (outside enrichment 0), and round 2 outside score 0 with probability 1. Greedy recommends B in round 1 (expected enrichment 10 vs 9.9 for A) and yields expected total enrichment 101.1, whereas recommending A in round 1 yields 109.9. Thus the theorem fails for non-stationary outside options. The fix is to add an explicit stationarity assumption (identical distribution of o_j(t) across t) to the model, or to prove a genuinely time-dependent characterization; as written, the central optimality claim needs qualification.
  2. [Section 5] The empirical claim of consistent outperformance is partly forced by the evaluation design. In Section 5, the synthetic data are generated from the paper's own model equations (enrichment and temptation drawn from anti-correlated multivariate normals, choices by choice-score maximization), and the baselines are precisely the estimators that ignore one or more of the model's components; such baselines are misspecified by construction, so the gap is expected. In Section 6, the MovieLens 'click data' are simulated from rating timestamps under a ratings-based recommender, the model is fit to those simulated data, and then the fitted model is treated as perfect information and used to evaluate the algorithms; this evaluates internal consistency of the estimation procedure rather than out-of-sample performance against real recommendation logs. The paper should refrain from claiming that the proposed algorithm 'consistently outperforms competitive baselines' in real-world settings; at most, the experiments demonstrate that the method works in the paper's own sandbox and can recover its model from simulated data. An evaluation with genuinely held-out behavioral data, or with synthetic data generated from a different structural model, would be needed to support the broader claim.
minor comments (5)
  1. [Section 3] In the statement of Theorem 1, 'maximizes the expected enrichment a receives' should read 'a user receives'; the sentence is missing a noun.
  2. [Section 4.2] The estimation loss in Algorithm 1 includes a term H(j,t) over the set S_j(t) union {o_j(t)} where the outside option choice value is drawn from N(mu, sigma). Since the realized outside option choice score is not observed, it would help to specify whether H(j,t) is an expectation over this draw or uses a sampled value in each SGD step; the pseudocode is ambiguous.
  3. [Section 5.1] The text says 'we report results for four distinct scenarios' but Figure 1 has three panels; the mapping from the four scenarios (enriching/tempting on-platform items crossed with perfect/partial information) to panels (a) and (b) should be stated explicitly in the caption.
  4. [Appendix B] The simulation parameter appendix uses 'a a normal distribution' in two places; this is a typo for 'a normal distribution'.
  5. [Section 2] The assumption that lambda^F_j > lambda^C_j is introduced without discussion of how it is enforced in estimation; the paper should note whether the optimizer is constrained to respect the inequality or whether it is only enforced by resampling during simulation.

Circularity Check

1 steps flagged · score 6.0 of 10

Empirical outperformance on 'real-world data' is simulated from the paper's own fitted model; Theorem 1 itself is independent.

  1. fitted input called prediction [Section 6, 'Estimating underlying model' and Section 6.1, 'Results']
    "We treat these estimated parameters as perfect information about the world the users and movies are in, which then allows us to observe the effects of recommendation algorithms on user consumption of movies. ... Using the estimated user-movie interaction model, we simulate 50 rounds of movie consumption. In each round, users are presented with a set of recommended movies from a specific algorithm and make choices based on estimated choice scores."

    The evaluation never observes real consumption; it simulates users with the same estimated choice scores, namely \hat C_j(i) = \hat\lambda^C_j \hat u_j(i) + (1 - \hat\lambda^C_j) \hat v_j(i), that the estimation framework fits and that the proposed greedy algorithm is designed to optimize. By Theorem 1, the proposed algorithm is the exact maximizer of expected consumed enrichment under this choice model, so its superiority over baselines that ignore temptation is a mathematical consequence of the fitted model rather than an empirical discovery. The only real data are used for fitting; the outcome variable is generated from the model's own equations, so the claimed 'real-world' outperformance reduces to the model's definitions.

full rationale

The paper's central theoretical contribution, Theorem 1, is a mathematical optimality claim about the stated model and is not circular: the proof in Appendix A is an exchange argument, and the greedy strategy is not defined in terms of the optimal value. No load-bearing self-citation or imported uniqueness theorem appears. The circularity lies in the empirical evaluation. In Section 6, the MovieLens data are used only to fit the model parameters; the subsequent 50 rounds of 'consumption' are simulated from the estimated choice scores rather than observed. Since the proposed recommender is, by Theorem 1, the exact optimizer of expected enrichment under that same fitted model, the reported outperformance over ratings-based and click-based baselines is a consequence of the model's assumptions, not an independent confirmation. The synthetic simulations in Section 5 are admittedly simulations generated from the model and illustrate the theorem rather than test it; they are not independent evidence, but they are not a fitted-input prediction in the same sense. A separate correctness concern, not a circularity, is that Theorem 1 as stated omits stationarity of the outside-option distribution: the proof's fixed ordering i_k presupposes a time-invariant distribution, and without it a non-stationary counterexample can beat greedy. That gap should be addressed under correctness review.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

The theoretical theorem only assumes the behavioral choice rule and stationary outside options. The estimation framework adds a low-rank factorization, a normal distribution for outside option choice scores, independence of outside option enrichment and choice score, and a survey-based obtainability of expected outside enrichment. The empirical claims additionally depend on generating data from the model itself, which is why the experimental evidence is not independent.

free parameters (5)
  • choice parameter lambda^C_j = estimated via SGD; mode ~0.25 in simulations
    User-specific weight on enrichment vs temptation when choosing; sampled from Beta(12.5, 37.5) in simulations, estimated in Section 4.2.
  • feedback parameter lambda^F_j = estimated via SGD; mode ~0.75 in simulations
    User-specific weight on enrichment vs temptation in ratings; assumed > lambda^C_j, estimated in Section 4.2.
  • latent vectors a_j, b_j, x_i, y_i = dimensionality d=3; all values estimated
    Low-rank factorization of enrichment and temptation in Section 4.1; all fitted to data via SGD.
  • outside option choice score parameters mu, sigma = estimated from data
    Assumed choice scores of outside options ~ N(mu, sigma^2); mu and sigma estimated in Section 4.2.
  • loss weights alpha, beta = not specified; alpha+beta=1
    Weights for rating loss vs click loss in Algorithm 1; concrete values not given.
assumptions (5)
  • domain assumption User choice is argmax of lambda^C_j u_j(i) + (1-lambda^C_j) v_j(i)
    Section 2 consumption stage; the core behavioral assumption, not derived.
  • ad hoc to paper lambda^F_j > lambda^C_j for every user
    Section 2 user feedback; asserted so ratings reflect enrichment more than choices do. Needed for the separation of the two attributes.
  • domain assumption Outside option choice scores follow N(mu, sigma^2) and are independent of outside option enrichment
    Section 4.2; needed for the closed-form expected enrichment and for identifiability of outside option parameters.
  • ad hoc to paper The distribution of the available outside option is stationary across rounds
    Implicit in Section 2 and used by the proof in Appendix A via a fixed greedy ordering i_1, i_2, ...; not stated as an assumption.
  • domain assumption Platform can ascertain E_l[u_j(o_l)] for each user, e.g., via surveys
    Section 4.2; load-bearing for the estimator, but not for the theoretical theorem.
invented entities (2)
  • Enrichment u_j(i) and temptation v_j(i) as separate latent item-user attributes
    purpose: To explain why users consume low-value high-appeal items and why ratings differ from choices
    These constructs are never measured directly; the paper's experiments generate them from the model. No real-world data distinguishes the two-attribute model from a single-utility model with noise.
  • Outside options o_j(t) with a distribution over user-specific alternatives
    purpose: To model the user's ability to leave the platform, which changes optimal recommendations
    Used everywhere in the model; the MovieLens experiment treats any out-of-sample movie as an outside option, but consumption is simulated, not observed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Recommendation and Temptation." pith.science (2026). https://pith.science/paper/BI4ZR3S6

@misc{pith2026241210595,
  author       = {Pith},
  title        = {Pith review of: Recommendation and Temptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BI4ZR3S6}},
  note         = {Machine review of arXiv:2412.10595}
}
read the original abstract

Traditional recommender systems based on revealed preferences often fail to capture the fundamental duality in user behavior, where consumption choices are driven by both inherent value (enrichment) and instant appeal (temptation). Consequently, these systems may generate recommendations that prioritize short-term engagement over long-lasting user satisfaction. We propose a novel recommender design that explicitly models the tension between enrichment and temptation. We introduce a behavioral model that accounts for how both enrichment and temptation influence user choices, while incorporating the reality of off-platform alternatives. Building on this model, we formulate a novel recommendation objective aligned with maximizing consumed enrichment and prove the optimality of a locally greedy recommendation strategy. Finally, we present an estimation framework that leverages the distinction between explicit user feedback and implicit choice data while making minimal assumptions about off-platform options. Through comprehensive evaluation using both synthetic simulations and real-world data from the MovieLens dataset, we demonstrate that our approach consistently outperforms competitive baselines that ignore temptation dynamics either by assuming revealed preferences or recommending solely based on enrichment. Our work represents a paradigm shift toward more nuanced and user-centric recommender design, with significant implications for developing responsible AI systems that genuinely serve users' long-term interests rather than merely maximizing engagement.

Figures

Figures reproduced from arXiv: 2412.10595 by the authors.

Figure 1
Figure 1. Overall individual enrichment provided by recommendation algorithms. (a) shows algorithm performance for enriching on-platform [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Frequency of consumption against enrichment and temp [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Overall individual enrichment provided by recommendation algorithms when enrichment and temptation of on-platform items have [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 20 canonical work pages

  1. [1]

    Arpit Agarwal, Nicolas Usunier, Alessandro Lazaric, and Maximilian Nickel

    work page

  2. [2]

    George A. Akerlof. 1991. Procrastination and Obedience. The American Economic Review 81, 2 (1991), 1–19. http://www.jstor.org/stable/2006817 RecSys ’25, September 22–26, 2025, Prague, Czech Republic Md Sanzeed Anwar, Paramveer S. Dhillon, and Grant Schoenebeck

    work page arXiv 1991

  3. [3]

    Dan Ariely and Klaus Wertenbroch. 2002. Procrastination, Deadlines, and Per- formance: Self-Control by Precommitment. Psychological Science 13, 3 (2002), 219–224. http://www.jstor.org/stable/40063710

    work page arXiv 2002

  4. [4]

    Allison J. B. Chaney, Brandon M. Stewart, and Barbara E. Engelhardt. 2018. How Algorithmic Confounding in Recommendation Systems Increases Homogeneity and Decreases Utility. InProceedings of the 12th ACM Conference on Recommender Systems (Vancouver, British Columbia, Canada) (RecSys ’18) . Association for Computing Machinery, New York, NY, USA, 224–232. h...

    arXiv 2018

  5. [5]

    Chi, and Minmin Chen

    Bo Chang, Alexandros Karatzoglou, Yuyan Wang, Can Xu, Ed H. Chi, and Minmin Chen. 2023. Latent User Intent Modeling for Sequential Recommenders. In Companion Proceedings of the ACM Web Conference 2023(Austin, TX, USA)(WWW ’23 Companion). Association for Computing Machinery, New York, NY, USA, 427–431. https://doi.org/10.1145/3543873.3584641

    arXiv 2023

  6. [6]

    Tom Cunningham, Sana Pandey, Leif Sigerson, Jonathan Stray, Jeff Allen, Bonnie Barrilleaux, Ravi Iyer, Smitha Milli, Mohit Kothari, and Behnam Rezaei. 2024. What We Know About Using Non-Engagement Signals in Content Ranking. arXiv:2402.06831 [cs.SI] https://arxiv.org/abs/2402.06831

    arXiv 2024

  7. [7]

    Ekstrand and Martijn C

    Michael D. Ekstrand and Martijn C. Willemsen. 2016. Behaviorism is Not Enough: Better Recommendations through Listening to Users. In Proceedings of the 10th ACM Conference on Recommender Systems (Boston, Massachusetts, USA) (RecSys ’16). Association for Computing Machinery, New York, NY, USA, 221–224. https: //doi.org/10.1145/2959100.2959179

    arXiv 2016

  8. [8]

    Daniel Fleder and Kartik Hosanagar. 2009. Blockbuster culture’s next rise or fall: The impact of recommender systems on sales diversity. Management Science 55, 5 (2009), 697–712

    work page 2009

Show all 39 references
  1. [9]

    Xavier Giné, Dean Karlan, and Jonathan Zinman. 2010. Put Your Money Where Your Butt Is: A Commitment Contract for Smoking Cessation.American Economic Journal: Applied Economics 2, 4 (2010), 213–235. http://www.jstor.org/stable/ 25760239

    2010

  2. [10]

    Faruk Gul and Wolfgang Pesendorfer. 2001. Temptation and Self-Control. Econo- metrica 69, 6 (2001), 1403–1435. http://www.jstor.org/stable/2692262

    2001

  3. [11]

    Faruk Gul and Wolfgang Pesendorfer. 2004. Self-Control and the Theory of Consumption. Econometrica 72, 1 (2004), 119–158. http://www.jstor.org/stable/ 3598852

    2004

  4. [12]

    Maxwell Harper and Joseph A

    F. Maxwell Harper and Joseph A. Konstan. 2015. The MovieLens Datasets: History and Context. ACM Trans. Interact. Intell. Syst. 5, 4, Article 19 (dec 2015), 19 pages. https://doi.org/10.1145/2827872

    2015 doi

  5. [13]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, YongDong Zhang, and Meng Wang. 2020. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation. In Proceedings of the 43rd International ACM SIGIR Confer- ence on Research and Development in Information Retrieva...

    2020

  6. [14]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural Collaborative Filtering. InProceedings of the 26th International Conference on World Wide Web(Perth, Australia) (WWW ’17). International World Wide Web Conferences Steering Committee, R...

    2017

  7. [15]

    Jon Kleinberg, Jens Ludwig, Sendhil Mullainathan, and Manish Ragha- van. 2024. The Inversion Problem: Why Algorithms Should Infer Men- tal State and Not Just Predict Behavior. Perspectives on Psychological Science 19, 5 (2024), 827–838. https://doi.org/10.1177/1745691623121213...

    2024 doi

  8. [16]

    Jon Kleinberg, Sendhil Mullainathan, and Manish Raghavan. 2022. The Challenge of Understanding What Users Want: Inconsistent Preferences and Engagement Optimization. In Proceedings of the 23rd ACM Conference on Economics and Com- putation (Boulder, CO, USA) (EC ’22). Associati...

    2022

  9. [17]

    David Laibson. 1997. Golden Eggs and Hyperbolic Discounting. The Quarterly Journal of Economics 112, 2 (1997), 443–477. http://www.jstor.org/stable/2951242

    1997

  10. [18]

    Jesse McCrosky and Brandi Geurkink. 2021. YouTube Regrets: A crowdsourced investigation into YouTube’s recommendation algorithm. Mozilla Foundation (2021)

    2021

  11. [19]

    Milkman, Todd Rogers, and Max H

    Katherine L. Milkman, Todd Rogers, and Max H. Bazerman. 2009. Highbrow Films Gather Dust: Time-Inconsistent Preferences and Online DVD Rentals. Management Science 55, 6 (2009), 1047–1059. https://doi.org/10.1287/mnsc.1080. 0994

    2009 doi

  12. [20]

    Smitha Milli, Luca Belli, and Moritz Hardt. 2021. From optimizing engagement to measuring value. In Proceedings of the 2021 ACM Conference on Fairness, Ac- countability, and Transparency. 714–722

    2021

  13. [21]

    Smitha Milli, Micah Carroll, Yike Wang, Sashrika Pandey, Sebastian Zhao, and Anca D. Dragan. 2023. Engagement, User Satisfaction, and the Amplification of Divisive Content on Social Media. arXiv:2305.16941 [cs.SI]

    2023 arXiv

  14. [22]

    Smitha Milli, Emma Pierson, and Nikhil Garg. 2023. Choosing the Right Weights: Balancing Value, Strategy, and Noise in Recommender Systems. arXiv:2305.17428 [cs.LG]

    2023 arXiv

  15. [23]

    Morewedge, Sendhil Mullainathan, Haaya F

    Carey K. Morewedge, Sendhil Mullainathan, Haaya F. Naushan, Cass R. Sunstein, Jon Kleinberg, Manish Raghavan, and Jens O. Ludwig. 2023. Human bias in algorithm design. Nature Human Behaviour 7, 11 (Nov. 2023), 1822–1824. https: //doi.org/10.1038/s41562-023-01724-4

    2023 doi

  16. [24]

    Nguyen, Pik-Mai Hui, F

    Tien T. Nguyen, Pik-Mai Hui, F. Maxwell Harper, Loren Terveen, and Joseph A. Konstan. 2014. Exploring the Filter Bubble: The Effect of Using Recommender Systems on Content Diversity. In Proceedings of the 23rd International Conference on World Wide Web (Seoul, Korea) (WWW ’14)...

    2014

  17. [25]

    Ted O’Donoghue and Matthew Rabin. 1999. Doing It Now or Later. American Economic Review 89, 1 (March 1999), 103–124. https://doi.org/10.1257/aer.89.1.103

    1999 doi

  18. [26]

    Kenny Peng, Manish Raghavan, Emma Pierson, Jon Kleinberg, and Nikhil Garg. 2023. Reconciling the accuracy-diversity trade-off in recommendations. arXiv:2307.15142 [cs.IR]

    2023 arXiv

  19. [27]

    Guy Raz. 2020. reCAPTCHA and Duolingo: Luis von Ahn. How I Built This with Guy Raz, NPR. https://podcasts.apple.com/us/podcast/recaptcha-and-duolingo- luis-von-ahn/id1150510297?i=1000475632807 Podcast episode, 1 hour 4 minutes

    2020

  20. [28]

    Recode Staff. 2017. Amazon? HBO? Netflix thinks its real competitor is... sleep. CNBC. https://www.cnbc.com/2017/04/17/amazon-hbo-netflix-thinks-its-real- competitor-is--sleep.html Netflix Q1 2017 earnings call

    2017

  21. [29]

    R. H. Strotz. 1955. Myopia and Inconsistency in Dynamic Utility Maximization. The Review of Economic Studies 23, 3 (12 1955), 165–180. https://doi.org/10.2307/ 2295722 arXiv:https://academic.oup.com/restud/article-pdf/23/3/165/6939058/23- 3-165.pdf

    1955

  22. [30]

    Sunstein and Richard H

    Cass R. Sunstein and Richard H. Thaler. 2003. Libertarian Paternalism Is Not an Oxymoron. The University of Chicago Law Review 70, 4 (2003), 1159–1202. http://www.jstor.org/stable/1600573

    2003

  23. [31]

    Thaler and Shlomo Benartzi

    Richard H. Thaler and Shlomo Benartzi. 2004. Save More Tomorrow ™: Using Behavioral Economics to Increase Employee Saving. Journal of Political Economy 112, S1 (2004), S164–S187. http://www.jstor.org/stable/10.1086/380085

    2004 doi

  24. [32]

    Thaler and H

    Richard H. Thaler and H. M. Shefrin. 1981. An Economic Theory of Self-Control. Journal of Political Economy 89, 2 (1981), 392–406. http://www.jstor.org/stable/ 1833317

    1981

  25. [33]

    Yaacov Trope and Ayelet Fishbach. 2000. Counteractive self-control in over- coming temptation. Journal of personality and social psychology 79, 4 (2000), 493

    2000

  26. [34]

    Luis von Ahn. 2023. How to make learning as addictive as social media. TED Talk. https://www.ted.com/talks/luis_von_ahn_how_to_make_learning_as_ addictive_as_social_media Accessed January 2025

    2023

  27. [35]

    Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural Graph Collaborative Filtering. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval (Paris, France) (SIGIR’19). Association for Computin...

    2019

  28. [36]

    Klaus Wertenbroch. 1998. Consumption Self-Control by Rationing Purchase Quantities of Virtue and Vice. Marketing Science 17, 4 (1998), 317–337. http: //www.jstor.org/stable/193203

    1998

  29. [37]

    Kaitlin Woolley and Riani Kenyon. 2023. I’ve been sucked in! The science of the rabbit hole. https://www.canvas8.com/library/reports/2023/02/03/ive-been- sucked-in-the-science-of-the-rabbit-hole Recommendation and Temptation RecSys ’25, September 22–26, 2025, Prague, Czech Rep...

    2023

  30. [39]

    As the figure shows, our algorithm maintains its superior perfor- mance, demonstrating its robustness to distributions of enrichment and temptation

  31. [2024]

    In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (Rio de Janeiro, Brazil) (FAccT ’24)

    System-2 Recommenders: Disentangling Utility and Engagement in Rec- ommendation Systems via Temporal Point-Processes. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (Rio de Janeiro, Brazil) (FAccT ’24). Association for Computing Machine...

    2024

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.