Pith. sign in

REVIEW 3 major objections 4 minor 41 references

This record's abstract claims calibrated recommenders let users switch preference groups as freely as traditional lists, and that outlier detection best reveals the underlying distributional structure.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

The record is internally inconsistent: the abstract promises a study of distribution structure in calibrated recommenders, but the full text is the previously published ENCODE CTR-modeling paper.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection The submission's abstract and full text are different papers: the calibrated-recommendation study described in the abstract is entirely absent, and the full text is a previously published IEEE TKDE paper by different authors. the 3 major comments →

arxiv 2508.13568 v1 pith:4KOISYTA submitted 2025-08-19 cs.IR cs.LG

Understanding Distribution Structure on Calibrated Recommendation Systems

Diego Correa da Silva , Denis Robson Dantas Boaventura , Mayki dos Santos Oliveira , Eduardo Ferreira da Silva , Joel Machado Pires , Frederico Ara\'ujo Dur\~ao This is my paper
classification cs.IR cs.LG
keywords calibrated recommendationgenre distributionoutlier detectionpreference groupsrecommendation diversitydistribution structureuser profilemovie recommendation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The abstract of this record argues that calibrated recommendation—forcing less-represented genres into the recommended list—does not cost users their freedom to change preference groups. It describes the user profile, the candidate item set, and the recommendation list as three G-dimensional genre distributions, and reports that fifteen models, especially outlier-detection models, can expose how those distributions are structured. If the finding held, calibration would gain a fairness and diversity benefit without the commonly feared side effect of locking users into their current tastes. However, the full text supplied with the record is a separate manuscript on long-term click-through-rate modeling; none of the fifteen models, three datasets, or results named in the abstract appear in that text, so the abstract's empirical claims are not supported by the presented body.

Core claim

The intended discovery is that calibrated recommendation lists behave like traditional recommendation lists with respect to how freely users move among preference groups, despite the extra constraint of including niche genres. In the paper's framing, each user's profile and each candidate or recommended list is a point in G-dimensional genre space, and the question is how these three distributions relate. The authors' intended contribution is a comparative diagnosis of that structure: outlier-detection models turn out to be the most informative tools, and the calibrated list is no more restrictive than the traditional list when a user's preference-group affiliations change. Because the attac

What carries the argument

The central object is the G-dimensional genre distribution: the user profile, the candidate item pool, and the recommendation list are each summarized as a vector over the system's G genres, making calibrated recommendation a comparison among three high-dimensional distributions. The paper's chosen instrument for that comparison is outlier detection, which scores how unusual one distribution is relative to another. This machinery is what supports the claim that calibrated lists occupy the same kind of distributional position as traditional lists and therefore preserve users' ability to change preference groups.

Load-bearing premise

The record's central claim stands on the premise that the supplied full text is the same study as the abstract, but it is a different manuscript on long-term click-through modeling; without that match the abstract's empirical findings have no presented evidence, even before asking whether a G-dimensional genre distribution plus outlier detection is the right way to measure preference-group mobility.

What would settle it

On three movie datasets, run a calibrated and a traditional recommender for the same users over consecutive recommendation rounds and record each user's preference-group membership before and after each round; if calibrated lists produce significantly fewer preference-group transitions than traditional lists, the 'same degree of freedom' claim is false. Separately, if outlier-detection scores on the recommendation-list distribution cannot distinguish calibrated from traditional lists, the claim that these models provide a better understanding of the structure fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Calibration can be treated as a safe diversity/fairness intervention: it need not reduce users' ability to shift from one preference group to another.
  • Outlier-detection scores on the recommendation-list distribution become a practical audit tool for detecting when a calibrated system is becoming too narrow.
  • The three-distribution comparison (profile, candidates, list) gives a standard coordinate system for comparing calibrated and traditional recommenders beyond item-level accuracy.
  • The same G-dimensional structure can be reused for other taxonomized item domains (music, news, video), not just movies.
  • Practitioners can use the 'same degree of preference-group change' claim as a measurable acceptance criterion when deploying calibrated recommenders.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural next step would be to define preference-group mobility quantitatively, for example as the transition probability between genre-distribution clusters across successive recommendation rounds, and test calibrated versus traditional lists on that metric.
  • The claim that outlier detection 'provides a better understanding' suggests an inexpensive auditing procedure: run off-the-shelf outlier scores on the list distribution to flag when calibration starts to restrict users, no retraining required.
  • Because the record's full text does not match its abstract, readers should treat the abstract's empirical findings as unverified until the fifteen models, three datasets, and results are located in a matching manuscript.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The submission, as identified by its abstract (arXiv:2508.13568), claims to study the distribution structure of calibrated recommendation systems. The abstract states that the authors implement fifteen models, evaluate on three movie-domain datasets, use G-dimensional genre distributions, find that outlier-detection models best reveal the structure, and conclude that calibrated recommendation lists allow users to shift preference groups to the same degree as traditional lists. However, the supplied full text is a completely different paper: ENCODE, a two-stage long-term user behavior modeling method for CTR prediction, with a header identifying it as arXiv:2508.13567v1. The body contains no mention of calibration, genre distributions, G-dimensional spaces, outlier detection, fifteen models, or the three movie datasets promised in the abstract. Consequently, the central claim of the abstract has no supporting implementation, experimental evidence, or derivation in the record under review.

Significance. If the abstract's claims were supported, the work could contribute to the evaluation of calibrated recommendation systems by proposing a distribution-structure analysis and arguing that calibrated lists preserve users' ability to change preference-group membership. That would be a useful result for the recommender-systems community. However, none of the claimed contributions is present in the supplied full text. The body text is a previously published IEEE TKDE paper on ENCODE, which has its own merits: it provides a detailed method, complexity analysis, and extensive offline experiments on industrial and public datasets, with ablations and hyperparameter studies. These strengths do not address the abstract's claim. The mismatch is not a local presentation issue; it removes the evidential basis for the paper's stated central result.

major comments (3)
  1. [Abstract vs. Full Text] The abstract's central claim has zero support in the supplied full text. The abstract describes fifteen models, three movie-domain datasets, G-dimensional genre distributions, outlier-detection models, and a conclusion about calibrated recommendation lists. The full text is the ENCODE paper on long-term CTR modeling, whose header reads 'arXiv:2508.13567v1' and whose content covers clustering-based interest extraction, metric-learning dimensionality reduction, and CTR prediction. No section defines the fifteen models, the three movie datasets, the distribution-structure analysis, or the calibration setting. This is a load-bearing mismatch: the record under review does not contain the study its own abstract describes.
  2. [Abstract, first paragraph] Even taken on its own, the abstract is under-specified in ways that prevent evaluation. The number of genres G is not defined; the three distributions (user profile, candidate items, recommendation list) are named but not formalized; the fifteen models are not enumerated; and the three movie datasets are not identified. Without these definitions, claims such as 'the models of outlier detection provide a better understanding of the structures' and 'the calibrated system creates recommendation lists that act similarly to traditional recommendation lists' cannot be checked. This is not a minor omission because these elements are the core of the claimed contribution.
  3. [Table 3 (body text)] Independently of the abstract mismatch, the full text's own evaluation contains a circularity concern. In Section 4.6.1 and Table 3, the Relevance Indicator (RI) is defined as the distance between each method's attention weights and the attention weights of DIN-L, with DIN-L's weights treated as ground truth. The paper then uses RI as evidence that ENCODE 'extracts more relevant long-term interest representations.' Since DIN-L is itself a model whose weights are not a verified ground truth, RI measures agreement with DIN-L, not objective relevance. The later claim that better RI correlates with better CTR performance is suggestive but does not resolve this definitional circularity.
minor comments (4)
  1. [Section 4.6.3] Heading typo: 'Our Rimensionality Reduction vs. SimHash' should read 'Our Dimensionality Reduction vs. SimHash'.
  2. [Author list] Spacing issue in the author list: 'Yunan Y e' should be 'Yunan Ye'.
  3. [Biographies] Typo in the biography of Xiaosong Yang: 'Bournemouth Unviersity' should be 'Bournemouth University'.
  4. [Table 3] The table title and caption call the metric 'RI' but the table has a column labeled 'CTR RI'; consider ensuring consistent notation. Also, DIN and DIN-L rows have no RI values, which should be explained in the caption.

Circularity Check

1 steps flagged

No direct circular derivation in the supplied body; the abstract's claims are unsupported by the mismatched full text, and the only definitional concern is the RI metric that measures relevance by proximity to DIN-L's attention weights.

specific steps
  1. self definitional [Section 4.6.1, Table 3]
    "Considering that DIN-L is recognized as the advanced accurate long-term sequence modeling algorithm, we regard the attention weights of DIN-L as ground truth and define Relevance Indicator (RI) as the distance between the output weights of each method and attention weights of DIN-L, to measure the relevance."

    The paper defines 'relevance' operationally as closeness to DIN-L's attention weights, then reports that ENCODE, which was designed to approximate target attention (the same mechanism DIN-L uses) with a consistent relevance metric, has the best RI. This is a definitional preference rather than an independent ground truth: the metric rewards methods that mimic DIN-L, and ENCODE is structurally close to DIN-L. However, this RI evidence is not the central performance claim; the paper's main CTR results are supported by independent AUC/GAUC comparisons against external baselines.

full rationale

The supplied full text is the ENCODE paper (arXiv:2508.13567v1) on long-term CTR modeling, while the abstract under review belongs to a different paper on calibrated recommendation systems (arXiv:2508.13568). The abstract's claims about fifteen models, three movie datasets, G-dimensional genre distributions, and outlier detection have no corresponding derivation or experimental evidence in the supplied body. This is a severe mismatch and a missing-support problem, but it is not itself a circular derivation. Within the ENCODE body, the main contribution—efficient two-stage interest modeling—is evaluated against external AUC and GAUC benchmarks, so the central results are not reductions to the paper's own inputs. The only definitionally circular element is the Relevance Indicator in Section 4.6.1, which defines ground-truth relevance weights as DIN-L's attention weights and then shows ENCODE is closest to them; because ENCODE was designed to approximate target attention like DIN-L, this metric is stacked by construction. Yet this RI analysis is secondary to the main independent benchmark results. Therefore, the circularity score is low: no load-bearing circularity in the core derivation, though the abstract/body mismatch prevents any circularity audit of the actual claimed study.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 0 invented entities

The record provides no free parameters for its stated central claim because the supporting study is absent; the only hyperparameters in the text (cluster number K=30, reduced dimension m=4, similarity scale beta=20, triplet margin alpha, auxiliary loss weight 0.1) belong to the ENCODE body, a different paper. No new entities are postulated by the abstract, and the body's cluster-based multi-interest representations are model components, not invented physical or conceptual entities.

axioms (3)
  • domain assumption A G-dimensional genre distribution sufficiently represents user profile, candidate items, and recommendation lists for calibration evaluation.
    Stated in the abstract; G, the distribution definitions, and the distance metric are never specified anywhere in the supplied text.
  • domain assumption Including less-represented genres via calibration is the right objective, and distribution-level comparison evaluates it.
    Abstract premise; the full text never mentions calibration, so the premise is never examined.
  • ad hoc to paper The supplied full text is the implementation and evaluation of the abstract's study.
    Required for any verification of the abstract's claims; contradicted on page 1, where the body is the ENCODE paper with header arXiv:2508.13567v1, received March 2024, accepted October 2024.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Understanding Distribution Structure on Calibrated Recommendation Systems." pith.science (2026). https://pith.science/paper/4KOISYTA

@misc{pith2026250813568,
  author       = {Pith},
  title        = {Pith review of: Understanding Distribution Structure on Calibrated Recommendation Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4KOISYTA}},
  note         = {Machine review of arXiv:2508.13568}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Traditional recommender systems aim to generate a recommendation list comprising the most relevant or similar items to the user's profile. These approaches can create recommendation lists that omit item genres from the less prominent areas of a user's profile, thereby undermining the user's experience. To solve this problem, the calibrated recommendation system provides a guarantee of including less representative areas in the recommended list. The calibrated context works with three distributions. The first is from the user's profile, the second is from the candidate items, and the last is from the recommendation list. These distributions are G-dimensional, where G is the total number of genres in the system. This high dimensionality requires a different evaluation method, considering that traditional recommenders operate in a one-dimensional data space. In this sense, we implement fifteen models that help to understand how these distributions are structured. We evaluate the users' patterns in three datasets from the movie domain. The results indicate that the models of outlier detection provide a better understanding of the structures. The calibrated system creates recommendation lists that act similarly to traditional recommendation lists, allowing users to change their groups of preferences to the same degree.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 40 canonical work pages

  1. [1]

    Practice on long sequential user behavior modeling for click-through rate prediction

    Qi Pi, Weijie Bian, Guorui Zhou, Xiaoqiang Zhu, and Kun Gai. Practice on long sequential user behavior modeling for click-through rate prediction. In ACM SIGKDD, pages 2671–2679, 2019

    work page 2019

  2. [2]

    Search- based user interest modeling with lifelong sequential behavior data for click-through rate prediction

    Qi Pi, Guorui Zhou, Yujing Zhang, Zhe Wang, Lejian Ren, Ying Fan, Xiaoqiang Zhu, and Kun Gai. Search- based user interest modeling with lifelong sequential behavior data for click-through rate prediction. InACM CIKM, pages 2685–2692, 2020

    work page 2020

  3. [3]

    User behavior retrieval for click- through rate prediction

    Jiarui Qin, Weinan Zhang, Xin Wu, Jiarui Jin, Yuchen Fang, and Yong Yu. User behavior retrieval for click- through rate prediction. In ACM SIGIR , pages 2347– 2356, 2020

    work page 2020

  4. [4]

    Sse-pt: Sequential recommendation via personal- ized transformer

    Liwei Wu, Shuqing Li, Cho-Jui Hsieh, and James Sharp- nack. Sse-pt: Sequential recommendation via personal- ized transformer. In ACM RecSys, pages 328–337, 2020

    work page 2020

  5. [5]

    Transformers4rec: Bridging the gap between nlp and sequential/session-based recommendation

    Gabriel de Souza Pereira Moreira, Sara Rabhi, Jeong Min Lee, Ronay Ak, and Even Oldridge. Transformers4rec: Bridging the gap between nlp and sequential/session-based recommendation. In ACM RecSys, pages 143–153, 2021

    work page 2021

  6. [6]

    Lifelong sequential modeling with personalized memorization for user response prediction

    Kan Ren, Jiarui Qin, Yuchen Fang, Weinan Zhang, Lei Zheng, Weijie Bian, Guorui Zhou, Jian Xu, Yong Yu, Xiaoqiang Zhu, and Kun Gai. Lifelong sequential modeling with personalized memorization for user response prediction. In ACM SIGIR , pages 565–574, 2019

    work page 2019

  7. [7]

    Large-scale modeling of mobile user click behaviors using deep learning

    Xin Zhou and Yang Li. Large-scale modeling of mobile user click behaviors using deep learning. In ACM RecSys, pages 473–483, 2021

    work page 2021

  8. [8]

    Contextual and sequential user em- beddings for large-scale music recommendation

    Casper Hansen, Christian Hansen, Lucas Maystre, Rishabh Mehrotra, Brian Brost, Federico Tomasi, and Mounia Lalmas. Contextual and sequential user em- beddings for large-scale music recommendation. In ACM RecSys, pages 53–62, 2020

    work page 2020

  9. [9]

    Sminet: State-aware multi-aspect interests representa- tion network for cold-start users recommendation

    Wanjie Tao, Yu Li, Liangyue Li, Zulong Chen, Hong Wen, Peilin Chen, Tingting Liang, and Quan Lu. Sminet: State-aware multi-aspect interests representa- tion network for cold-start users recommendation. In AAAI, pages 8476–8484, 2022

    work page 2022

  10. [10]

    Contextual-bandit based personalized recommen- dation with time-varying user interests

    Xiao Xu, Fang Dong, Yanghua Li, Shaojian He, and Xin Li. Contextual-bandit based personalized recommen- dation with time-varying user interests. In AAAI, pages 6518–6525, 2020

    work page 2020

  11. [11]

    Deep learning for click-through rate estimation

    Weinan Zhang, Jiarui Qin, Wei Guo, Ruiming Tang, and Xiuqiang He. Deep learning for click-through rate estimation. In IJCAI, 2021. JOURNAL OF IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING,VOL. ,NO. ,APRIL 2024 12

    work page 2021

  12. [12]

    Lifelong sequential modeling with personalized memorization for user response pre- diction

    Kan Ren, Jiarui Qin, Yuchen Fang, Weinan Zhang, Lei Zheng, Weijie Bian, Guorui Zhou, Jian Xu, Yong Yu, Xiaoqiang Zhu, et al. Lifelong sequential modeling with personalized memorization for user response pre- diction. In ACM SIGIR, pages 565–574, 2019

    work page 2019

  13. [13]

    Sampling is all you need on modeling long-term user behaviors for CTR prediction

    Yue Cao, Xiaojiang Zhou, Jiaqi Feng, Peihao Huang, Yao Xiao, Dayao Chen, and Sheng Chen. Sampling is all you need on modeling long-term user behaviors for CTR prediction. In ACM CIKM, pages 2974–2983, 2022

    work page 2022

  14. [14]

    Factorization machines

    Steffen Rendle. Factorization machines. In IEEE ICDM, pages 995–1000, 2010

    work page 2010

  15. [15]

    Field- weighted factorization machines for click-through rate prediction in display advertising

    Junwei Pan, Jian Xu, Alfonso Lobos Ruiz, Wenliang Zhao, Shengjun Pan, Yu Sun, and Quan Lu. Field- weighted factorization machines for click-through rate prediction in display advertising. In WWW, pages 1349–1357, 2018

    work page 2018

  16. [16]

    Wide & deep learning for recommender systems

    Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen An- derson, Greg Corrado, Wei Chai, Mustafa Ispir, Rohan Anil, Zakaria Haque, Lichan Hong, Vihan Jain, Xiaob- ing Liu, and Hemal Shah. Wide & deep learning for recommender systems. In ACM RecSys, 2016

    work page 2016

  17. [17]

    Attentional factorization machines: Learning the weight of feature interactions via attention networks

    Jun Xiao, Hao Ye, Xiangnan He, Hanwang Zhang, Fei Wu, and Tat-Seng Chua. Attentional factorization machines: Learning the weight of feature interactions via attention networks. In IJCAI, pages 3119–3125, 2017

    work page 2017

  18. [18]

    Deepfm: A factorization-machine based neural network for CTR prediction

    Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. Deepfm: A factorization-machine based neural network for CTR prediction. In IJCAI, pages 1725–1731, 2017

    work page 2017

  19. [19]

    Product-based neural networks for user response prediction

    Yanru Qu, Han Cai, Kan Ren, Weinan Zhang, Yong Yu, Ying Wen, and Jun Wang. Product-based neural networks for user response prediction. In IEEE ICDM, pages 1149–1154, 2016

    work page 2016

  20. [20]

    xdeepfm: Combining explicit and implicit feature in- teractions for recommender systems

    Jianxun Lian, Xiaohuan Zhou, Fuzheng Zhang, Zhongxia Chen, Xing Xie, and Guangzhong Sun. xdeepfm: Combining explicit and implicit feature in- teractions for recommender systems. In ACM SIGKDD, pages 1754–1763, 2018

    work page 2018

  21. [21]

    Deep & cross network for ad click predictions

    Ruoxi Wang, Bin Fu, Gang Fu, and Mingliang Wang. Deep & cross network for ad click predictions. In ACM SIGKDD, pages 12:1–12:7, 2017

    work page 2017

  22. [22]

    Deep neural networks for youtube recommendations

    Paul Covington, Jay Adams, and Emre Sargin. Deep neural networks for youtube recommendations. In ACM RecSys, pages 191–198, 2016

    work page 2016

  23. [23]

    Session-based recommen- dations with recurrent neural networks

    Bal ´azs Hidasi, Alexandros Karatzoglou, Linas Bal- trunas, and Domonkos Tikk. Session-based recommen- dations with recurrent neural networks. In ICLR, 2016

    work page 2016

  24. [24]

    Deep inter- est evolution network for click-through rate prediction

    Guorui Zhou, Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, and Kun Gai. Deep inter- est evolution network for click-through rate prediction. In AAAI, pages 5941–5948, 2019

    work page 2019

  25. [25]

    Deep session interest network for click-through rate prediction

    Yufei Feng, Fuyu Lv, Weichen Shen, Menghan Wang, Fei Sun, Yu Zhu, and Keping Yang. Deep session interest network for click-through rate prediction. In IJCAI, pages 2301–2307, 2019

    work page 2019

  26. [26]

    Bert4rec: Sequential rec- ommendation with bidirectional encoder representa- tions from transformer

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang. Bert4rec: Sequential rec- ommendation with bidirectional encoder representa- tions from transformer. In ACM CIKM , pages 1441– 1450, 2019

    work page 2019

  27. [27]

    Deep interest network for click-through rate prediction

    Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. Deep interest network for click-through rate prediction. In ACM SIGIR, pages 1059–1068, 2018

    work page 2018

  28. [28]

    Multi-interest network with dynamic routing for recommendation at tmall

    Chao Li, Zhiyuan Liu, Mengmeng Wu, Yuchi Xu, Huan Zhao, Pipei Huang, Guoliang Kang, Qiwei Chen, Wei Li, and Dik Lun Lee. Multi-interest network with dynamic routing for recommendation at tmall. In ACM CIKM, pages 2615–2623, 2019

    work page 2019

  29. [29]

    What aspect do you like: Multi-scale time-aware user interest modeling for micro-video recommendation

    Hao Jiang, Wenjie Wang, Yinwei Wei, Zan Gao, Yin- glong Wang, and Liqiang Nie. What aspect do you like: Multi-scale time-aware user interest modeling for micro-video recommendation. In ACM MM , pages 3487–3495, 2020

    work page 2020

  30. [30]

    Controllable multi-interest framework for recommendation

    Yukuo Cen, Jianwei Zhang, Xu Zou, Chang Zhou, Hongxia Yang, and Jie Tang. Controllable multi-interest framework for recommendation. In ACM SIGKDD , pages 2942–2951, 2020

    work page 2020

  31. [31]

    Efficient long sequential user data modeling for click-through rate prediction

    Qiwei Chen, Yue Xu, Changhua Pei, Shanshan Lv, Tao Zhuang, and Junfeng Ge. Efficient long sequential user data modeling for click-through rate prediction. arXiv, abs/2209.12212, 2022

    Pith/arXiv arXiv 2022

  32. [32]

    Twin: Two-stage interest network for lifelong user behavior modeling in ctr prediction at kuaishou

    Jianxin Chang, Chenbin Zhang, Zhiyi Fu, Xiaoxue Zang, Lin Guan, Jing Lu, Yiqun Hui, Dewei Leng, Yanan Niu, Yang Song, et al. Twin: Two-stage interest network for lifelong user behavior modeling in ctr prediction at kuaishou. In ACM SIGKDD, pages 3785– 3794, 2023

    work page 2023

  33. [33]

    Razenshteyn, and Ludwig Schmidt

    Alexandr Andoni, Piotr Indyk, Thijs Laarhoven, Ilya P . Razenshteyn, and Ludwig Schmidt. Practical and opti- mal LSH for angular distance. In NeurIPS, pages 1225– 1233, 2015

    work page 2015

  34. [34]

    Similarity estimation techniques from rounding algorithms

    Moses S Charikar. Similarity estimation techniques from rounding algorithms. In ACM STOC, pages 380– 388, 2002

    work page 2002

  35. [35]

    Distance metric learning with application to clustering with side-information

    Eric Xing, Michael Jordan, Stuart J Russell, and An- drew Ng. Distance metric learning with application to clustering with side-information. In NeurIPS, 2002

    work page 2002

  36. [36]

    Facenet: A unified embedding for face recog- nition and clustering

    Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recog- nition and clustering. In CVPR, pages 815–823, 2015

    work page 2015

  37. [37]

    McAuley, Christopher Targett, Qinfeng Shi, and Anton van den Hengel

    Julian J. McAuley, Christopher Targett, Qinfeng Shi, and Anton van den Hengel. Image-based recommen- dations on styles and substitutes. In ACM SIGIR, pages 43–52, 2015

    work page 2015

  38. [38]

    Sparse attentive memory network for click-through rate prediction with long sequences

    Qianying Lin, Wen-Ji Zhou, Yanshi Wang, Qing Da, Qing-Guo Chen, and Bing Wang. Sparse attentive memory network for click-through rate prediction with long sequences. In ACM CIKM, pages 3312–3321, 2022

    work page 2022

  39. [39]

    The movie- lens datasets: History and context

    F Maxwell Harper and Joseph A Konstan. The movie- lens datasets: History and context. ACM Transactions on Interactive Intelligent Systems (TIIS), 5(4):1–19, 2015

    work page 2015

  40. [40]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In arXiv, 2014

    work page 2014

  41. [41]

    Improved deep metric learning with multi-class n-pair loss objective

    Kihyuk Sohn. Improved deep metric learning with multi-class n-pair loss objective. NeurIPS, 2016. JOURNAL OF IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING,VOL. ,NO. ,APRIL 2024 13 Wen-Ji Zhou received the BSc and MSc de- gree in computer science from Nanjing Univer- sity (NJU), Nanjing, China, in 2016 and 2019, respectively. He is currently a senior...

    work page 2016

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.