Pith. sign in

REVIEW 4 major objections 6 minor 29 references

PinEqualizer: Full Funnel Content Exploration and Debiasing System at Pinterest

T0 review · 4 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A system that debiases every stage of the recommendation funnel, from corpus selection to final ranking, is credited with a 350% increase in fresh-content impressions at Pinterest, along with long-term engagement and creator-diversity gains

desk verdict A credible industry-scale full-funnel cold-start system paper with real deployment evidence; the headline causal numbers are softer than they look, but the component launches and measurement framework make it worth a serious referee. read the letter →

arxiv 2607.22518 v1 pith:BHRSJ6DY submitted 2026-07-24 cs.IR cs.LG

Olafur Gudmundsson , Bo Zhao , Huayi Liao , Anna Kiyantseva , Sai Xiao , Heath Vinicombe , Mostafa Keikha , Luke DeLuccia
show 10 more authors
Zihao Chen Junpeng Hou Weijie Jiang Bhawna Juneja Andreanne Lemay Wei-Ting Lin Keyvan Moghadam Jiaxing Qu Zhiqing Rao Zhihua Zhang
This is my paper · ORCID
classification cs.IRcs.LG
keywords cold-startrecommendersystemscontentexplorationdebiasingfull-funnelUCBecosystemmeasurementframework
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 the content cold-start problem in large-scale search and recommendation systems is not solved by exploration alone: the entire multi-stage funnel is biased toward existing content, so new content must be debiased and given fair access at every stage. The authors built and deployed PinEqualizer across Pinterest's Homefeed, Related Pins, and Search surfaces, intervening in corpus selection, retrieval, ranking, and the final utility layer. They report that this full-funnel approach increased fresh-content impressions by 350%, grew the graduated fresh-content corpus by 41%, and nearly doubled the number of successful content providers, while also lifting user engagement. A core part of the contribution is a three-layer measurement framework—a long-term fresh-content holdout, a content graduation metric, and a fast under-explored engagement volume metric—that lets the team validate long-term value while iterating quickly. If correct, the paper shows that reducing bias against fresh content is a concrete, measurable way to improve both engagement and content ecosystem health at industry scale.

What carries the argument

The central machinery is the full-funnel exploration-and-debiasing pipeline, held together by a three-layer measurement framework. At corpus selection, a dedicated exploration corpus scores fresh items by a posterior engagement estimate that combines a model-based prior with observed engagement, retiring items once they graduate or are deemed low-engaging. At retrieval, debiasing comes from dedicated exploration channels, weighted random walks over the pin-board graph, content-only embeddings, and unified learned retrieval models. At ranking, the paper uses engagement-feature dropout, feature imputation, content-type-aware calibration, regularization, and training-data augmentation to reduce

What would settle it

Run the fresh-content holdout with a matched control that removes an equivalent volume of existing content chosen by predicted engagement rate rather than uniformly at random; if the engagement delta shrinks toward zero, the incremental value attributed to fresh content is an artifact of removing low-value old content. Separately, log content IDs in both A/B arms and measure whether under-explored content that graduates in the treatment arm subsequently appears as under-explored engagement in the control arm; nonzero cross-arm traffic would directly test the paper's assertion that leakage is m

Watch

Extended reading notes

Core claim

The central claim is that cold-start content fails in production not mainly because it is low quality, but because every stage of the funnel—corpus selection, retrieval, ranking, and utility—contains bias favoring older, more connected content. The paper shows that countering this bias with a coordinated set of interventions, rather than relying on explicit exploration alone, yields large and lasting gains. Concretely, the deployed system produced a 350% increase in fresh-content impressions, a 41% year-over-year increase in content that graduated within 28 days, double the number of successful content providers, and substantial session-level engagement gains measured by a fresh-content hold

Load-bearing premise

The headline gains rest on the fresh-content holdout assumption that randomly removing an equivalent volume of existing content is engagement-neutral; if the removed content would have earned engagement anyway, the measured lift is not a clean estimate of the system's causal contribution, and a second premise—that content-level leakage in the under-explored engagement metric is minimal—is asserted rather than measured.

Editorial extensions

If this is right

  • If the holdout result is valid, freshly created and explored content provides incremental engagement beyond an equivalent volume of existing content, justifying continued investment in exploration rather than purely exploiting the current corpus.
  • Reducing bias across the funnel lowers the need for high-volume explicit exploration, which in turn reduces the short-term engagement tradeoff typically associated with exploration mechanisms.
  • The under-explored engagement volume metric enables fast, user-segmented A/B experiments that track long-term content value, allowing engineering teams to iterate on exploration improvements without waiting for long-term holdout results.
  • The full-funnel bottleneck analysis provides a practical sequencing rule: fix upstream corpus and retrieval constraints before investing heavily in ranking-stage exploration, since any single stage can throttle fresh-content distribution.
  • Search-specific safeguards, such as relevance weighting and minimum relevance thresholds on UCB bonuses, show that exploration can be applied on relevance-sensitive surfaces without degrading query-to-result relevance.

Reading between the lines

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

  • The especially large shopping-session lift (5.3x the non-shopping gain in North America) suggests that catalog-driven content, which arrives in bulk and lacks graph connectivity, is disproportionately dependent on fresh-content exploration; other platforms with similar merchant uploads may see comparable effects.
  • The paper's argument that content-level leakage under-estimates its under-explored engagement gains is plausible but unmeasured; a direct cross-arm leakage log would either confirm that the reported gains are conservative or require revising them downward.
  • The sequencing lesson—fix corpus and retrieval before ranking—implies that many production systems could be leaving ranking-stage exploration improvements on the table simply because upstream funnel stages are still throttling fresh content.
  • The finding that individual dropout of engagement features outperforms uniform dropout is a concrete, transferable design choice; testing it across other deep ranking architectures would show whether it generalizes beyond Pinterest's setup.
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

4 major / 6 minor

Summary. The paper describes PinEqualizer, a full-funnel content exploration and debiasing system deployed at Pinterest across Homefeed, Related Pins, and Search. The system spans corpus selection, retrieval, ranking, and utility layers, combining dedicated exploration corpora, debiased retrieval and ranking features, and explicit UCB-style exploration. A three-layer measurement framework is proposed: a long-term fresh-content holdout as a north-star metric, a content graduation corpus metric as an intermediate proxy, and an under-explored content engagement volume metric for short-term user-segmented A/B experiments. The authors report large site-wide gains, including +24%/+49% YoY incremental session gains from the fresh-content holdout, +41% growth in the graduated content corpus, +37%/+13%/+27% cumulative gains in under-explored engagement across surfaces, +99% growth in successful content providers, and claim a 350% increase in fresh-content impressions since 2024.

Significance. If the reported results are valid, this is a significant industry-scale demonstration that a coordinated, whole-funnel approach to cold-start can improve fresh-content distribution, user engagement, and ecosystem health simultaneously. The paper's measurement framework, which attempts to link a long-term holdout to intermediate corpus metrics and short-term experiment metrics, is a useful practical contribution. The component-level A/B results in Table 2 provide rare transparency about which interventions work, and the candid discussion of the Search-surface failure mode (strong regularization degrading relevance) and the decision to prefer a simpler heuristic over neural linear UCB add credibility. However, the published manuscript contains a formula error in the core UCB definition, and the causal interpretation of the north-star holdout is not fully supported by the evidence presented. These issues are fixable but require substantive revision.

major comments (4)
  1. [§4.4, Eq. (2)] As printed, Eq. (2) reads UCB_i = α√(1+β·impressions_i), which grows with impression count. This contradicts the stated intent that 'fewer impressions mean higher uncertainty' and would cause the exploration bonus to increase for already heavily served content, the opposite of exploration. The formula should be inverse (e.g., α/√(1+β·impressions_i)). This is a load-bearing algorithmic definition; please correct the equation and ensure the surrounding text and comparisons in Table 2 are consistent.
  2. [§2.1, Fig. 1a, Table 1] The fresh-content holdout randomly removes an 'equivalent amount' of content from the production group by content-ID hash, but the paper does not demonstrate that the randomly removed content is engagement-neutral relative to the fresh content ablated in the treatment. If the random removal disproportionately removes high- or low-engagement Pins, the observed engagement delta is a biased estimate of the incremental value of fresh content. Additionally, the YoY comparisons in Table 1 are not isolated from concurrent changes (e.g., content-supply composition, other ranking launches, seasonality); no difference-in-differences, synthetic control, or control surface is provided. The paper should report the engagement distribution of the removed set versus the ablated fresh set and provide a robustness check or explicitly state the limitation.
  3. [§1 and Table 1] The headline claim of a '350% increase in fresh content impressions' is prominently stated in the introduction and abstract but is not backed by any metric definition, table, or methodology in the paper. Please define the fresh-content impression metric (e.g., share of impressions from content younger than N days), specify the baseline period and the measurement window, or remove the claim from the high-level summary if it cannot be substantiated.
  4. [§5.2, Table 2] The component-level A/B lifts are described as 'simply aggregating the gains from individual launches' but no aggregation rule is given. It is unclear whether the percentages are additive across non-overlapping experiments, compounded, or adjusted for overlap, and how the 'strict guardrail on overall engagement tradeoff' was enforced. Without this information, the summed lifts in Table 2 are hard to interpret. Please state the aggregation methodology and note how overlapping or sequential launches are handled.
minor comments (6)
  1. [§2.2, §2.3] The thresholds X, Z, and the graduation window Y are reported as empirically chosen, but no sensitivity analysis or confidence information is provided. For a metric intended as a robust intermediate proxy, a short robustness discussion would be valuable.
  2. [§4.1, Eq. (1), and §5] The symbol N is overloaded: it denotes prior strength in Eq. (1), a reference batch size in the text around Eq. (4), and the fresh-content age cutoff in Table 2. Please use distinct notation (e.g., N_prior, N_ref, N_days) to avoid confusion.
  3. [§5.2] Typo: 'Neural Liner Bandit' should be 'Neural Linear Bandit'.
  4. [§4.4, Eq. (2)] The equation appears with an extra rendering artifact ('√︁'); please ensure the final typeset equation is clean.
  5. [§2.3] The claim that content-level leakage is 'minimal' is supported by two qualitative observations. A measurement of leakage or a more formal argument would strengthen the validity of the under-explored engagement metric.
  6. [§5.1] The definition of 'successful content providers' is given only as 'content's site-wide engagement volume share above a certain threshold' without the threshold value. Please provide the threshold or a range.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central measurement rests on an external session-level fresh-content holdout; intermediate defined metrics and one non-load-bearing self-citation do not make the derivation circular.

full rationale

The paper's central causal claim is anchored by the fresh-content holdout in §2.1/Fig 1a and Table 1, which is an external user-session benchmark: treatment removes fresh content while control removes an equivalent volume by content-ID hash, and the delta is measured in successful sessions, not in any metric the system optimizes. None of the reported lifts is produced by fitting a parameter and then predicting the same data: the graduation threshold X is a data-informed definition, but the reported graduation growth is a measured corpus statistic, not a model output fit to it. Under-explored engagement volume is the system's direct objective, but it is measured in user-segmented A/B experiments and, as §2.3 acknowledges, content-level leakage would attenuate rather than inflate it; its link to graduation is definitional (graduation means accumulating X positive engagements), but this is an explicitly labeled proxy chain, with the north-star holdout providing independent validation. The only same-team citation, [7], is used for implementation details of training data augmentation and regularization and is not load-bearing for any central claim. Remaining weaknesses—engagement-neutrality of the random-removal control, YoY confounds, and the apparent sign inconsistency in Eq. (2)—are internal-validity or correctness concerns, not circularity.

Assumptions & free parameters 11 free parameters · 7 assumptions · 3 invented entities

The paper introduces no physical entities; the invented constructs here are measurement definitions and a system component. The larger burden sits in free parameters: essentially all algorithmic constants (prior strength, exploration strength, decay, regularization, thresholds X/Y/Z, provider threshold, fresh-content window) are tuned or data-derived and undisclosed, so the reported magnitudes cannot be reproduced from the text. The key domain assumptions are the validity of the equivalent-volume holdout control and the minimality of leakage in the under-explored metric; both are acknowledged by the authors only partially.

free parameters (11)
  • Prior strength N (Eq 1) = not reported
    Controls how strongly the model-based prior α blends with empirical engagement in the posterior score r̂; tuned in production; magnitude unstated.
  • Graduation threshold X = not reported; data-derived
    'We identified specific cumulative engagement threshold X based on our empirical data where the variance of engagement rate decreases significantly' (§2.2). Defines the key intermediate success metric.
  • Graduation window Y = 28 days in Table 1
    Choice of window for the graduation metric; Table 1 uses 28-day graduation; no sensitivity analysis reported.
  • Low-engagement threshold Z = not reported
    Defines 'explored low-engaging content' as fully-explored content whose engagement-rate upper confidence bound is below Z (§2.2); hand-chosen, value undisclosed.
  • Exploration strength α (Eq 2) = not reported
    UCB bonus scale; tuned per surface and not reported.
  • Impression-decay β (Eq 2) = not reported
    Controls how quickly the UCB bonus decays with accumulated impressions; tuned, value undisclosed.
  • Reference batch size N0 (α scaling) = not reported
    Chosen to keep exploration magnitude stable as step-(2) sample size N varies in Neural Linear UCB (§4.4).
  • Ridge λ (Eq 4) and relevance exponent γ = not reported
    Neural-Linear-UCB regularization and the search-surface relevance scaling UCB' = relevance^γ · UCB; both tuned and undisclosed.
  • Successful-content-provider threshold = not reported
    Defines the +99% YoY metric as providers 'with their content's site-wide engagement volume share above a certain threshold' (§5.1); threshold never specified.
  • Fresh-content definition N days = not stated
    Table 2 reports 'all fresh (created < N days)'; N is never defined.
  • Thompson sampling prior α and dropout rates = not reported
    Beta prior on provider engagement (§4.1) and the two engagement-feature dropout strategies evaluated in §4.2; rates undisclosed.
assumptions (7)
  • domain assumption Beta-Binomial engagement-rate model (posterior Beta(α+e, n−e))
    §4.1 assumes each item's engagement rate is a static latent with binomial observation noise; standard bandit modeling, but engagement rates of fresh content may drift over time.
  • domain assumption Random equivalent-volume ablation is a valid control for candidate-pool size
    §2.1/Fig 1a: the north-star holdout delta is interpreted as the incremental value of fresh content only if randomly removing the same volume of content from the production cohort is engagement-neutral. This is also the weakest assumption of the paper.
  • domain assumption Content-level leakage in the under-explored engagement metric is minimal
    §2.3: the paper asserts leakage causes under-estimation of gains and gives two reasons it is small; load-bearing for all Table 2 experiment metrics, and flagged by the authors as an approximation rather than a measured quantity.
  • domain assumption Variance-drop in engagement rate identifies a meaningful graduation point
    §2.2: the graduation threshold X is selected where engagement-rate variance 'decreases significantly'; the proxy validity of the graduation metric rests on this empirical pattern.
  • domain assumption The relevance model has lower fresh-content bias than the engagement model
    §4.4 'Role of Relevance in Search Utility': justifies boosting relevance weight for fresh content and scaling UCB by relevance^γ in search; stated without supporting analysis.
  • standard math Neural-Linear-UCB linear surrogacy (logit ≈ φ(x)ᵀθ)
    §4.4, Eq (3): from [22]; the paper avoids computing θ and uses the pretrained binary-prediction logit as a cheap surrogate for the mean reward.
  • domain assumption Impression count is a valid uncertainty proxy (fewer impressions ⇒ higher uncertainty)
    §4.4, Eq (2): the heuristic UCB identifies uncertainty with impression count; note this assumption is in tension with the printed equation, which increases with impressions.
invented entities (3)
  • Content graduation metric independent evidence
    purpose: Intermediate proxy between short-term experiment metrics and the long-term holdout: newly ingested content that accumulates X positive engagements within Y days is 'graduated' from exploration to exploitation.
    A new measurement construct rather than a physical entity; its validity is checked against the fresh-content holdout north star, and graduation growth is claimed to correlate with holdout session gains.
  • Under-explored content ledger entry independent evidence
    purpose: Defines the short-term experiment metric (positive engagement volume on under-explored content) and the membership rule for the exploration corpus: content that is neither graduated nor explored-low-engaging.
    A new corpus classification; falsifiable through its claimed correlation with graduation and with the holdout's user-level engagement outcomes.
  • Dedicated exploration corpus independent evidence
    purpose: Corpus-selection stage component: a curated subset of high-potential fresh items selected by Thompson sampling/posterior score and retired on graduation or low upper-confidence bound.
    A new system component, included for completeness; its contribution is individually A/B-tested (+7.75% engagement, +16.92% under-explored engagement in Table 2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of PinEqualizer: Full Funnel Content Exploration and Debiasing System at Pinterest." pith.science (2026). https://pith.science/paper/BHRSJ6DY

@misc{pith2026260722518,
  author       = {Pith},
  title        = {Pith review of: PinEqualizer: Full Funnel Content Exploration and Debiasing System at Pinterest},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BHRSJ6DY}},
  note         = {Machine review of arXiv:2607.22518}
}
read the original abstract

In this paper, we propose a new solution for addressing the content cold-start problem in industry-scale search and recommender systems. Compared to prior approaches, we have made the following new contributions: 1) our solution spans the entire multi-stage funnel and generalizes well for both search and recommendation surfaces, 2) our solution reduces bias favoring existing content, allowing more accurate model prediction across content types and reducing short-term tradeoffs associated with high volumes of explicit content exploration, 3) our solution is evaluated with a scalable measurement framework that enables fast short-term experimentation while validating long-term impact. We have iteratively built and successfully deployed this new system at Pinterest in the past two years and observed significant improvements in fresh content exploration, overall user engagement, and content ecosystem health.

Figures

Figures reproduced from arXiv: 2607.22518 by the authors.

Figure 1
Figure 1. a) Fair fresh content holdout experiment setup, hold￾out group (treatment) removes fresh content, production group (control) removes same amount of content randomly. b) The long￾term fresh content holdout demonstrates that the short-term costs of exploration are outweighed by long-term retention benefits. c) Content graduation denotes the threshold at which an item’s en￾gagement value goes from uncertain to predicta… view at source ↗
Figure 2
Figure 2. System Diagram: Full Funnel Content Exploration and Debiasing [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 1 canonical work pages

  1. [1]

    Peter Auer, Nicolò Cesa-Bianchi, and Paul Fischer. 2002. Finite-time analysis of the multiarmed bandit problem.Machine learning47, 2 (2002), 235–256

    2002

  2. [3]

    Josh Beal, Eric Kim, Jinfeng Rao, Rex Wu, Dmitry Kislyuk, and Charles Rosenberg

  3. [4]

    Chi, and Minmin Chen

    Bo Chang, Changping Meng, He Ma, Shuo Chang, Yang Gu, Yajun Peng, Jingchen Feng, Yaping Zhang, Shuchao Bi, Ed H. Chi, and Minmin Chen. 2024. Cluster Anchor Regularization to Alleviate Popularity Bias in Recommender Systems. In Companion Proceedings of the ACM Web Conference 2024(Singapore, Singapore) (WWW ’24). Association for Computing Machinery, New Yor...

    arXiv 2024

  4. [5]

    Gaode Chen, Ruina Sun, Yinjie Jiang, Tianxiang Li, Yinlong Dai, Qifan Shi, Xiaojie Qin, Jialong Fu, Piaoyi Chen, Ronggeng Huang, Na Li, Qi Zhang, Jiang Liang, Han Li, and Kun Gai. 2025. A Cold-start Recommendation System at Kuaishou Designed from the Short-video Perspective. InCompanion Proceedings of the ACM on Web Conference 2025(Sydney NSW, Australia)(...

    doi:10.1145/3701716 2025

  5. [6]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep Neural Networks for YouTube Recommendations. InProceedings of the 10th ACM Conference on Recommender Systems(Boston, Massachusetts, USA)(RecSys ’16). Association for Computing Machinery, New York, NY, USA, 191–198. doi:10.1145/2959100. 2959190

    doi:10.1145/2959100 2016

  6. [7]

    Saeed Ebrahimi, Weijie Jiang, Jaewon Yang, Olafur Gudmundsson, Yucheng Tu, and Huizhong Duan. 2026. Warmer for Less: A Cost-Efficient Strategy for Cold- Start Recommendations at Pinterest. InProceedings of the ACM Web Conference 2026(United Arab Emirates)(WWW ’26). Association for Computing Machinery, New York, NY, USA, 8105–8114. doi:10.1145/3774904.3792824

    arXiv 2026

  7. [8]

    Chantat Eksombatchai, Pranav Jindal, Jerry Zitao Liu, Yuchen Liu, Rahul Sharma, Charles Sugnet, Mark Ulrich, and Jure Leskovec. 2018. Pixie: A System for Recommending 3+ Billion Items to 200+ Million Users in Real-Time. InProceedings of the 2018 World Wide Web Conference(Lyon, France)(WWW ’18). International World Wide Web Conferences Steering Committee, ...

    arXiv 2018

  8. [9]

    Schapire

    Lihong Li, Wei Chu, John Langford, and Robert E. Schapire. 2010. A contextual- bandit approach to personalized news article recommendation. InProceedings of the 19th International Conference on World Wide Web(Raleigh, North Carolina, USA)(WWW ’10). Association for Computing Machinery, New York, NY, USA, 661–670. doi:10.1145/1772690.1772758

    arXiv 2010

Show all 29 references
  1. [10]

    Yunze Luo, Yuezihan Jiang, Yinjie Jiang, Gaode Chen, Jingchi Wang, Kaigui Bian, Peiyi Li, and Qi Zhang. 2025. Online Item Cold-Start Recommendation with Popularity-Aware Meta-Learning. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.1(T...

    2025

  2. [11]

    Nikil Pancha, Andrew Zhai, Jure Leskovec, and Charles Rosenberg. 2022. Pinner- Former: Sequence Modeling for User Representation at Pinterest. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (Washington DC, USA)(KDD ’22). Association for ...

    2022

  3. [12]

    Deepak Kumar Panda and Sanjog Ray. 2022. Approaches and algorithms to mitigate cold start problems in recommender systems: a systematic literature review.J. Intell. Inf. Syst.59, 2 (Oct. 2022), 341–366. doi:10.1007/s10844-022- 00698-5

    2022 doi

  4. [13]

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Tran, Jonah Samost, et al

  5. [14]

    Tran, Jonah Samost, Maciej Kula, Ed H

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Keshavan, Trung Vu, Lukasz Heidt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Maheswaran Sathiamoorthy. 2023. Recommender systems with generative retrieval. InProceedings of the 37th Inter...

    2023

  6. [15]

    Schein, Alexandrin Popescul, Lyle H

    Andrew I. Schein, Alexandrin Popescul, Lyle H. Ungar, and David M. Pennock

  7. [16]

    Anima Singh, Trung Vu, Nikhil Mehta, Raghunandan Keshavan, Maheswaran Sathiamoorthy, Yilin Zheng, Lichan Hong, Lukasz Heldt, Li Wei, Devansh Tandon, Ed Chi, and Xinyang Yi. 2024. Better Generalization with Semantic IDs: A Case Study in Ranking for Recommendations. InProceeding...

    2024 doi

  8. [17]

    Chi, Cristos Goodrow, Su-Lin Wu, Lexi Baugher, and Minmin Chen

    Yi Su, Xiangyu Wang, Elaine Ya Le, Liang Liu, Yuening Li, Haokai Lu, Benjamin Lipshitz, Sriraj Badam, Lukasz Heldt, Shuchao Bi, Ed H. Chi, Cristos Goodrow, Su-Lin Wu, Lexi Baugher, and Minmin Chen. 2024. Long-Term Value of Ex- ploration: Measurements, Findings and Algorithms. ...

    2024

  9. [18]

    Maksims Volkovs, Guangwei Yu, and Tomi Poutanen. 2017. DropoutNet: address- ing cold start in recommender systems. InProceedings of the 31st International Conference on Neural Information Processing Systems(Long Beach, California, USA)(NIPS’17). Curran Associates Inc., Red Hoo...

    2017

  10. [19]

    Chi, Cristos J

    Jianling Wang, Haokai Lu, Sai Zhang, Bart Locanthi, Haoting Wang, Dylan Greaves, Benjamin Lipshitz, Sriraj Badam, Ed H. Chi, Cristos J. Goodrow, Su-Lin Wu, Lexi Baugher, and Minmin Chen. 2023. Fresh Content Needs More Attention: Multi-funnel Fresh Content Recommendation. InPro...

    2023

  11. [20]

    Ruoxi Wang, Rakesh Shivanna, Derek Cheng, Sagar Jain, Dong Lin, Lichan Hong, and Ed Chi. 2021. DCN V2: Improved Deep & Cross Network and Practical Lessons for Web-scale Learning to Rank Systems. InProceedings of the Web Conference 2021(Ljubljana, Slovenia)(WWW ’21). Associatio...

    2021

  12. [21]

    Xue Xia, Pong Eksombatchai, Nikil Pancha, Dhruvil Deven Badani, Po-Wei Wang, Neng Gu, Saurabh Vishwas Joshi, Nazanin Farahpour, Zhiyuan Zhang, and An- drew Zhai. 2023. TransAct: Transformer-based Realtime User Action Model for Recommendation at Pinterest. InProceedings of the ...

    2023

  13. [22]

    Pan Xu, Zheng Wen, Handong Zhao, and Quanquan Gu. 2020. Neu- ral Contextual Bandits with Deep Representation and Shallow Exploration. arXiv:2012.01780 [cs.LG] https://arxiv.org/abs/2012.01780

    2020 arXiv

  14. [23]

    Hamilton, and Jure Leskovec

    Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L. Hamilton, and Jure Leskovec. 2018. Graph Convolutional Neural Networks for Web-Scale Recommender Systems. InProceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining(London...

    2018

  15. [24]

    Andrew Zhai, Hao-Yu Wu, Eric Tzeng, Dong Huk Park, and Charles Rosenberg

  16. [25]

    Weitong Zhang, Dongruo Zhou, Lihong Li, and Quanquan Gu. 2021. Neural Thompson Sampling. arXiv:2010.00827 [cs.LG] https://arxiv.org/abs/2010.00827

    2021 arXiv

  17. [26]

    Dongruo Zhou, Lihong Li, and Quanquan Gu. 2020. Neural contextual bandits with UCB-based exploration. InProceedings of the 37th International Conference on Machine Learning (ICML’20). JMLR.org, Online, Article 1065, 11 pages

    2020

  18. [2002]

    InProceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval(Tampere, Finland)(SIGIR ’02)

    Methods and metrics for cold-start recommendations. InProceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval(Tampere, Finland)(SIGIR ’02). Association for Computing Machinery, New York, NY, USA, 253–260. doi:10.11...

  19. [2019]

    InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining(Anchorage, AK, USA)(KDD ’19)

    Learning a Unified Embedding for Visual Search at Pinterest. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining(Anchorage, AK, USA)(KDD ’19). Association for Computing Machinery, New York, NY, USA, 2412–2420. doi:10.1145/3292500.3330739

  20. [2023]

    Recommender systems with generative retrieval.Advances in Neural Information Processing Systems36 (2023), 10299–10315

    2023

  21. [2026]

    arXiv:2603.03544 [cs.CV] https://arxiv.org/abs/2603.03544

    PinCLIP: Large-scale Foundational Multimodal Representation at Pinterest. arXiv:2603.03544 [cs.CV] https://arxiv.org/abs/2603.03544

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.