Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

A Hybrid Cross-Stage Coordination Pre-ranking Model for Online Recommendation Systems

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that training the pre-ranking stage on unexposed candidates from the whole retrieval-to-ranking stream, with a margin-adjusted contrastive loss, improves long-tail precision and reports 14.9% higher conversions in live…

desk verdict Solid industrial result with a real online lift, but the appendix proof for Margin InfoNCE has a sign error and does not establish the claimed potential-positive selection mechanism. read the letter →

arxiv 2502.10284 v1 pith:C5CV2P3E submitted 2025-02-14 cs.IR cs.AI

Binglei Zhao , Houying Qi , Guang Xu , Mian Ma , Xiwei Zhao , Feng Mei , Sulong Xu , Jinghe Hu This is my paper · ORCID
classification cs.IRcs.AI
keywords pre-rankingcross-stagecoordinationsampleselectionbiaslong-tailprecisioncontrastivelearningMarginInfoNCEmulti-stagerecommendationconsistency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that the pre-ranking stage of a cascading recommender—the cheap first-pass model that narrows millions of retrieved candidates before the full ranking stage—should be trained on the whole candidate stream instead of only on items the ranking stage exposed. It introduces a hybrid sample-construction scheme that gathers four levels of unexposed data (ranking leftovers, pre-ranking leftovers, in-batch negatives, and pool-sampled negatives) and a hybrid objective that combines listwise order-consistency losses with a Margin InfoNCE contrastive loss. The central claim is that this cross-stage coordination improves long-tail precision, lets pre-ranking act as a bridge between retrieval and ranking, and mitigates sample selection bias without a meaningful latency increase. On a large e-commerce production dataset and in a seven-day online test, the method reports up to 14.9% higher conversion rate and 1.3% higher click rate.

What carries the argument

The load-bearing mechanism is the Margin InfoNCE loss: a contrastive loss over positives, hard negatives (unexposed ranking and pre-ranking candidates), and easy negatives (in-batch and pool-sampled negatives), with an additive margin $m$ applied to hard-negative cosine similarities so the model treats hard negatives more gently than easy ones. The appendix supports this with a gradient comparison showing that the margin loss reduces the gradient on a hard negative as its score nears zero, whereas the BCE gradient stays larger. The supporting machinery is hybrid sample construction—non-uniform sampling and rearrangement of the ranking sequence into a 'ground truth' order, plus listwise in-batch sampling—and the global (ListMLE) and local consistency losses that set up the order signal the margin loss refines.

What would settle it

During online serving, log the follow-up clicks and purchases of unexposed items that HCCP scores highly; if their positive-feedback rate is no higher than the baseline's high-scoring items, or if removing the margin from the loss leaves that rate unchanged, the claim that Margin InfoNCE selects potential positives is unsupported.

Watch

Extended reading notes

Core claim

The central discovery is that unexposed candidates from the upstream and downstream stages, which are normally discarded or treated as uniform negatives, can be reorganized into a guiding training signal for pre-ranking. The paper's Margin InfoNCE loss separates hard negatives (likely relevant but unexposed candidates) from easy negatives by subtracting a margin from the cosine similarity of hard negatives, and the appendix argues that this makes the loss assign smaller gradients to hard negatives near score zero than standard cross-entropy does, slowing their descent and preserving potential positives. Combined with global and local listwise consistency losses, the design produces a pre-ranker that outperforms consistency-only baselines across top and long-tail metrics. The paper reports that, deployed on a production e-commerce homepage, it contributed up to 14.9% UCVR and 1.3% UCTR gains.

Load-bearing premise

The argument depends on the premise that the unexposed candidates set aside by ranking and pre-ranking still contain enough genuine potential positives for the margin heuristic to save; if they are mostly true negatives, the contrastive term mainly injects noise into pre-ranking learning.

Editorial extensions

If this is right

  • Pre-ranking can be trained on data drawn from the entire retrieval-to-ranking stream, not just exposed impressions, with serving latency increasing by only about 0.73% at the 99th percentile.
  • Treating unexposed ranking and pre-ranking candidates as hard negatives and in-batch or pool-sampled negatives as easy negatives improves hit rates at both small and large K values, which correspond to top and long-tail items.
  • Consistency-only pre-ranking methods improve order alignment with the ranking stage but leave long-tail bias largely unsolved; adding long-tail precision optimization is what drives the reported gains.
  • List-wise organization of the multi-level samples keeps offline training complexity manageable even when the data volume grows from hundreds of millions to billions of instances.

Reading between the lines

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

  • The same hard/easy margin contrastive treatment could plausibly transfer to the ranking stage itself, where retrieved-but-unexposed items are also discarded, although the paper does not test this.
  • The much larger conversion gain than click gain suggests long-tail items surfaced by the method convert at a disproportionately high rate; a per-item conversion-lift analysis would test whether that is the mechanism.
  • The non-uniform sampling threshold is calibrated from the observed crossover where pre-ranking's click hitrate exceeds the ranking model's; other systems would need to re-derive this threshold from their own data.
  • Because the public dataset lacks unexposed stream data, the paper's full claims are validated only where such logs exist; an independent check would need a public benchmark that includes unexposed candidates.
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

5 major / 5 minor

Summary. This paper proposes HCCP, a pre-ranking model for e-commerce recommendation that constructs hybrid training samples from the whole cascade (ranking sequence, pre-ranking sequence, in-batch negatives, and pool negatives) and jointly optimizes listwise consistency losses with a Margin InfoNCE contrastive loss. Evaluation is reported on the Taobao dataset, a proprietary JD production dataset, and a 7-day JD online A/B test, with claimed gains of up to 14.9% UCVR and 1.3% UCTR. An appendix presents a gradient-based argument intended to show that Margin InfoNCE selects potential positives from hard negatives.

Significance. If the reported online gains are reliable, the empirical contribution is practically significant for industrial pre-ranking systems: the paper shows a plausible path to training pre-ranking on unexposed data from upstream and downstream stages with only 0.73% added TP99 latency. The paper ships a TensorFlow-style pseudocode for the core loss, a decomposition of contributions via ablations, and a complexity table, all of which help reproducibility. The main weakness is that the theoretical justification for the central mechanism, the Margin InfoNCE loss, is not established by the appendix argument, and the experimental reporting lacks the statistical detail needed to judge the offline and online comparisons. The online A/B lift provides independent support for the central empirical claim, but the mechanism claim needs substantial revision.

major comments (5)
  1. [Appendix A] The proof of the central mechanism contains a sign error and does not establish the stated conclusion. Starting from Eq. (5), with phi(theta)=cos(theta)-m, the hard-negative logit is y_i - m*v_i. Factoring e^{-m*v_j/tau} out of numerator and denominator gives the hard-negative exponent (y_n + m*(v_j - v_n))/tau, not (y_n - m*(v_j - v_n))/tau as written in Eq. (9); Eq. (10) inherits the error. More importantly, the comparison f(y_n)<g(y_n) near y_n=0 shows only that the loss slows down the suppression of every hard negative, not that potential positives are discriminated from true hard negatives. The concluding sentence of the appendix, claiming that the loss 'thereby effectively selects the potential positives,' is therefore unsupported. A valid argument would need to show that high-scoring hard negatives receive larger relative score gains under Margin InfoNCE than under the baseline loss, or the paper should provide direct empirical evidence on the N2-N3 candidate pool.
  2. [Section 3.2.1] The rearrangement rule r_c = r - (e + l_click)*N_r is introduced and described as producing the 'optimal guiding ground truth' for pre-ranking, but no derivation or ablation is provided to justify optimality. The global consistency loss in Eq. (3) and the local consistency loss in Eq. (4) are trained on this rearranged order, so the contribution of the cross-stage coordination idea is entangled with this specific priority rule. The paper should either prove optimality under a stated criterion or compare against plausible alternatives, such as using the raw ranking order or exposure-first order without click priority, in the ablation study.
  3. [Sections 3.3.2 and 4.1.4] The exact loss used in training is ambiguous. Eq. (5) defines a single Margin InfoNCE loss, but Section 4.1.4 states that the final weighted score of the Three Tower is optimized on N2-N3 samples while user-item embedding dot-products are optimized on N2-N5 samples, with the final loss being the average of the two. Algorithm 1 additionally introduces a stabilized logit l_update = (logits_m + beta*logits)/(1+beta) with beta annealed from 9999 downward, which does not appear in Eq. (5). The paper needs to state, for each negative group, which loss variant is used and whether Eq. (5) or Algorithm 1 is the production objective; otherwise the ablation in Table 1 comparing Margin InfoNCE with plain InfoNCE cannot be reproduced.
  4. [Sections 4.2 and 4.3] The empirical claims are reported without estimates of uncertainty. Table 1 and Table 3 give percentage-point differences over a single validation day with no confidence intervals, significance tests, or repeated runs, and the differences between some ablations are small (e.g., HCCP(w/o Margin) versus HCCP(Ours) on ISH@1000). The online A/B section similarly omits traffic split, number of users or requests, confidence intervals, and the definition of the percentage lifts. These omissions weaken both the offline and online evidence, and the paper should report at least the standard errors or significance levels for the key comparisons.
  5. [Section 4.2.1] The Taobao evaluation does not test the method's distinctive components. Section 4.1.1 says the Taobao dataset lacks multi-level unexposed data, which creates a significant gap for Rethink and for HCCP, but Table 2 then reports HCCP on Taobao without specifying which variant was used. If HCCP on Taobao reduces to consistency training on impressions, the result should be labeled accordingly and the claim that 'HCCP achieves a slight improvement' should be scoped to that configuration.
minor comments (5)
  1. [Appendix A] In Eq. (9), the definition 'B_h = B_r ∪ B_e is the hard negative set' is inconsistent with the earlier definition B_e = B_i ∪ B_c as the easy negative set; the hard set should be B_r ∪ B_p.
  2. [Section 3.1.1] The calibration loss in Eq. (1) is normalized by |U| but not by the sequence length |R_t|; if the user-wise average is intended, the notation should say so explicitly, because otherwise the loss scale depends on the list length.
  3. [Section 4.1.4] The paper states that performance is sensitive to hyperparameters such as the global consistency weight and the margin value, but no sensitivity analysis is shown; adding a small table or figure would make this statement verifiable.
  4. [Algorithm 1] The beta annealing schedule in Algorithm 1 is unexplained, and the relationship between the pseudocode and Eq. (5) should be clarified, since the pseudocode appears to implement a different interpolation between margin-adjusted and original logits.
  5. [Throughout] There are several typos: 'Expect N1' should be 'Except N1' in Section 3.3.2, 'Futher' should be 'Further' in Section 4.4, and 'reachs' should be 'reaches' in Section 4.4.3.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the empirical claims rest on held-out and online A/B evidence, and the appendix proof gap is a rigor issue, not a definitional identity.

full rationale

The paper's central empirical claims are supported by held-out JD production data and a seven-day online A/B test, so the reported UCVR and UCTR gains are genuine predictions, not fitted parameters renamed as results. The Margin InfoNCE loss combines InfoNCE with an additive margin borrowed from AM-Softmax; the margin and temperature are hyperparameters tuned through ablation, but the comparison to HCCP(w/o Margin) and the independent online deployment provide external evidence for the overall method. The only place with a claimed first-principles result is the appendix, which attempts to prove that the margin softens gradients on hard negatives. That mathematical claim, even if flawed, is not circular: the proof does not define the loss in terms of the conclusion, and the conclusion is not an input to the loss. The appendix has a likely algebraic sign error in Eq. 9 and a logical gap between 'slower descent on all hard negatives' and 'selecting potential positives,' but these are correctness and rigor concerns, not cases where an output is equivalent to an input by construction. The only author-overlapping citation is reference [26], where Jinghe Hu is a co-author of both the cited paper and this paper; it is used in Related Work to acknowledge prior consistency-oriented pre-ranking work and is not load-bearing for HCCP's design or evaluation. No load-bearing self-citation chain, imported uniqueness theorem, ansatz smuggled by citation, or renamed fitted result was found.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The method depends on several user-chosen hyperparameters and on assumptions about unexposed data that are not independently verified. The lists above capture the load-bearing choices.

free parameters (8)
  • margin m in Margin InfoNCE = 0.9
    Set via ablation; controls how much harder hard negatives are pushed than the positive. The appendix proof depends on this value.
  • temperature tau = 0.1
    Set via ablation; controls softmax sharpness in the contrastive loss.
  • global consistency loss weight lambda_c = 0.05
    Chosen through ablation; the paper states performance is sensitive to this weight.
  • task loss weights lambda_t (click/purchase/diversity) = 0.98, 0.2, 0.1
    Manually set weights for multi-task calibration losses.
  • local consistency weight alpha = 0.05
    Weight for the ListMLE loss on the exposure sequence.
  • margin InfoNCE weight beta = 0.5
    Weight for the negative contrastive loss.
  • ranking sequence chunk sample rates = 0.5, 0.125, 0.1, 0.05, 0.025, 0.01, 0.005
    Non-uniform sampling rates from top to tail chunks; hand-designed based on the ASH@K analysis.
  • sampled sequence sizes (R_c, B_p, pool negatives) = 35, 15, 10
    Chosen per request for training; affects data volume and difficulty.
assumptions (4)
  • domain assumption Unclicked and unexposed items can be used as negatives in contrastive learning without introducing bias that outweighs the benefit.
    Section 3.3.2 states user feedback on N2-N5 is unknown; the method assumes contrastive objectives improve discriminative capacity despite this.
  • domain assumption The ranking model's predictions on top items are accurate enough to serve as soft ground truth for consistency learning.
    Section 3.3.1 aligns pre-ranking to ranking under the assumption that top ranks are correct.
  • ad hoc to paper The rearrangement rule r_c = r - (e + l_click) * N_r yields the optimal guiding order for pre-ranking.
    Section 3.2.1 introduces this rule with intuitive justification but no theoretical or empirical validation.
  • domain assumption Hard negatives N2-N3 contain a non-negligible density of potential positive items.
    The loss is designed to protect potential positives among hard negatives; the appendix argues this heuristically but does not measure it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Hybrid Cross-Stage Coordination Pre-ranking Model for Online Recommendation Systems." pith.science (2026). https://pith.science/paper/C5CV2P3E

@misc{pith2026250210284,
  author       = {Pith},
  title        = {Pith review of: A Hybrid Cross-Stage Coordination Pre-ranking Model for Online Recommendation Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C5CV2P3E}},
  note         = {Machine review of arXiv:2502.10284}
}
read the original abstract

Large-scale recommendation systems often adopt cascading architecture consisting of retrieval, pre-ranking, ranking, and re-ranking stages. With strict latency requirements, pre-ranking utilizes lightweight models to perform a preliminary selection from massive retrieved candidates. However, recent works focus solely on improving consistency with ranking, relying exclusively on downstream stages. Since downstream input is derived from the pre-ranking output, they will exacerbate the sample selection bias (SSB) issue and Matthew effect, leading to sub-optimal results. To address the limitation, we propose a novel Hybrid Cross-Stage Coordination Pre-ranking model (HCCP) to integrate information from upstream (retrieval) and downstream (ranking, re-ranking) stages. Specifically, cross-stage coordination refers to the pre-ranking's adaptability to the entire stream and the role of serving as a more effective bridge between upstream and downstream. HCCP consists of Hybrid Sample Construction and Hybrid Objective Optimization. Hybrid sample construction captures multi-level unexposed data from the entire stream and rearranges them to become the optimal guiding "ground truth" for pre-ranking learning. Hybrid objective optimization contains the joint optimization of consistency and long-tail precision through our proposed Margin InfoNCE loss. It is specifically designed to learn from such hybrid unexposed samples, improving the overall performance and mitigating the SSB issue. The appendix describes a proof of the efficacy of the proposed loss in selecting potential positives. Extensive offline and online experiments indicate that HCCP outperforms SOTA methods by improving cross-stage coordination. It contributes up to 14.9% UCVR and 1.3% UCTR in the JD E-commerce recommendation system. Concerning code privacy, we provide a pseudocode for reference.

Figures

Figures reproduced from arXiv: 2502.10284 by the authors.

Figure 1
Figure 1. Cascade architecture in recommendation system. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Architecture of HCCP based on a three-tower model. (A) Hybrid sample construction module to obtain multi-level [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Ranking Sequence Construction. (A) Non-uniform [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: List-wise in-batch sampling. Top and tail items [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: CTR and CVR improvements of items with differ [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Relative ISH/ISPH@100 of HCCP, HCCP(w/o Up), [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: 𝑓 (𝑦𝑛) and 𝑔(𝑦𝑛) functions. 𝑣𝑗 = ∥uj ∥ ∥ij ∥ , 𝑐𝑜𝑠𝜃 𝑗 = uj · ij ∥uj ∥ ∥ij ∥ , 𝜙𝑗 = 𝑐𝑜𝑠𝜃 𝑗 − 𝑚 , where 𝑗 and 𝑛(𝑛 ∈ 𝐵 ℎ ) are the positive and negative items to be calculated the partial derivative. 𝐵 ℎ = 𝐵 𝑟 ∪ 𝐵 𝑒 is the hard negative set and 𝐵 𝑒 = 𝐵 𝑖∪𝐵 𝑐 is the easy n…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. EGA-V1: Unifying Online Advertising with End-to-End Learning

    cs.IR 2025-05 conditional novelty 6.0 of 10

    EGA-V1 unifies advertising ranking and auction into a single non-autoregressive generative model with cluster attention, and is reported to beat multi-stage cascades on Meituan's ad traffic.

Reference graph

Works this paper leans on

39 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [1]

    Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Greg Hullender. 2005. Learning to rank using gradient descent. InProceedings of the 22nd International Conference on Machine Learning (Bonn, Germany) (ICML ’05). Association for Computing Machinery, New York, NY, USA, 89–96. https: //doi.org/10.1145/1102351.1102363

    arXiv 2005

  2. [2]

    Chris J.C. Burges. 2010. From RankNet to LambdaRank to LambdaMART: An Overview. Technical Report MSR-TR-2010-82. https://www.microsoft.com/en- us/research/publication/from-ranknet-to-lambdarank-to-lambdamart-an- overview/

    work page 2010

  3. [3]

    Zhe Cao, Tao Qin, Tie-Yan Liu, Ming-Feng Tsai, and Hang Li. 2007. Learning to rank: from pairwise approach to listwise approach. In Proceedings of the 24th International Conference on Machine Learning (Corvalis, Oregon, USA) (ICML ’07). Association for Computing Machinery, New York, NY, USA, 129–136. https: //doi.org/10.1145/1273496.1273513

    arXiv 2007

  4. [4]

    Fredric C. Gey. 1994. Inferring probability of relevance using the method of logistic regression. In Proceedings of the 17th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval (Dublin, Ireland) (SIGIR ’94). Springer-Verlag, Berlin, Heidelberg, 222–231

    work page 1994

  5. [5]

    Daniel Gillick, Sayali Kulkarni, Larry Lansing, Alessandro Presta, Jason Baldridge, Eugene Ie, and Diego Garcia-Olano. 2019. Learning Dense Representations for Entity Retrieval. arXiv:1909.10506 [cs.CL]

    arXiv 2019

  6. [6]

    Sheng, Keng Yeow Tay, Walter Romano, and Shuo Li

    Bin Gu, Victor S. Sheng, Keng Yeow Tay, Walter Romano, and Shuo Li. 2015. Incremental Support Vector Learning for Ordinal Regression. IEEE Transactions on Neural Networks and Learning Systems 26, 7 (2015), 1403–1416. https://doi. org/10.1109/TNNLS.2014.2342533

    work page arXiv 2015

  7. [7]

    Siyu Gu and Xiangrong Sheng. 2022. On Ranking Consistency of Pre-ranking Stage. arXiv:2205.01289 [cs.IR]

    arXiv 2022

  8. [8]

    Xiaoqiang Gui, Yueyao Cheng, Xiang-Rong Sheng, Yunfeng Zhao, Guoxian Yu, Shuguang Han, Yuning Jiang, Jian Xu, and Bo Zheng. 2023. Calibration- compatible Listwise Distillation of Privileged Features for CTR Prediction. arXiv:2312.08727 [cs.IR]

    work page Pith review arXiv 2023

Show all 39 references
  1. [9]

    Po-Sen Huang, Xiaodong He, Jianfeng Gao, Li Deng, Alex Acero, and Larry Heck. 2013. Learning Deep Structured Semantic Models for Web Search Using Clickthrough Data. In Proceedings of the 22nd ACM International Conference on Information & Knowledge Management (San Francisco, Ca...

    2013

  2. [10]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense Passage Retrieval for Open- Domain Question Answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP...

    2020 doi

  3. [11]

    Jingjing Li, Ke Lu, Zi Huang, and Heng Tao Shen. 2017. Two Birds One Stone: On both Cold-Start and Long-Tail Recommendation. In Proceedings of the 25th ACM International Conference on Multimedia (Mountain View, California, USA) (MM ’17). Association for Computing Machinery, Ne...

    2017

  4. [12]

    Longbin Li, Chao Zhang, Sen Li, Yun Zhong, Qingwen Liu, and Xiaoyi Zeng

  5. [13]

    Xiangyang Li, Bo Chen, Huifeng Guo, Jingjie Li, Chenxu Zhu, Xiang Long, Su- jian Li, Yichao Wang, Wei Guo, Longxia Mao, Jinxing Liu, Zhenhua Dong, and Ruiming Tang. 2022. IntTower: The Next Generation of Two-Tower Model for Pre-Ranking System. In Proceedings of the 31st ACM In...

    2022

  6. [14]

    Siyi Liu and Yujia Zheng. 2020. Long-tail Session-based Recommendation. In Proceedings of the 14th ACM Conference on Recommender Systems (Virtual Event, Brazil) (RecSys ’20). Association for Computing Machinery, New York, NY, USA, 509–514. https://doi.org/10.1145/3383313.3412222

    2020

  7. [15]

    Tie-Yan Liu. 2009. Learning to Rank for Information Retrieval. Found. Trends Inf. Retr. 3, 3 (mar 2009), 225–331. https://doi.org/10.1561/1500000016 WWW Companion ’25, April 28-May 2, 2025, Sydney, NSW, Australia. Binglei Zhao, et al. Algorithm 1 A Tensorflow-style Pseudocode ...

    2009 doi

  8. [16]

    Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. 2017. Sphereface: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition . 212–220

    2017

  9. [17]

    Weiyang Liu, Yandong Wen, Zhiding Yu, and Meng Yang. 2016. Large-margin softmax loss for convolutional neural networks. arXiv preprint arXiv:1612.02295 (2016)

    2016 arXiv

  10. [19]

    Xu Ma, Pengjie Wang, Hui Zhao, Shaoguo Liu, Chuhan Zhao, Wei Lin, Kuang- Chih Lee, Jian Xu, and Bo Zheng. 2021. Towards a Better Tradeoff between Effectiveness and Efficiency in Pre-Ranking: A Learnable Feature Selection based Approach. In Proceedings of the 44th International...

    2021

  11. [20]

    Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Distributed representations of words and phrases and their compositionality. In Proceedings of the 26th International Conference on Neural Information Processing Systems - Volume 2 (Lake Tahoe, Neva...

    2013

  12. [21]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)

    2018 arXiv

  13. [22]

    Palangi, L

    H. Palangi, L. Deng, Y. Shen, J. Gao, X. He, J. Chen, X. Song, and R. Ward. 2015. Semantic Modelling with Long-Short-Term Memory for Information Retrieval. arXiv:1412.6629 [cs.IR]

    2015 arXiv

  14. [23]

    Haojie Pan, Chengyu Wang, Minghui Qiu, Yichang Zhang, Yaliang Li, and Jun Huang. 2020. Meta-KD: A meta knowledge distillation framework for language model compression across domains. arXiv preprint arXiv:2012.01266 (2020)

    2020 arXiv

  15. [24]

    Jiarui Qin, Jiachen Zhu, Bo Chen, Zhirong Liu, Weiwen Liu, Ruiming Tang, Rui Zhang, Yong Yu, and Weinan Zhang. 2022. RankFlow: Joint Optimization of Multi- Stage Cascade Ranking Systems as Flows. In Proceedings of the 45th International ACM SIGIR Conference on Research and Dev...

    2022

  16. [26]

    Jinbo Song, Ruoran Huang, Xinyang Wang, Wei Huang, Qian Yu, Mingming Chen, Yafei Yao, Chaosheng Fan, Changping Peng, Zhangang Lin, Jinghe Hu, and Jingping Shao. 2022. Rethinking Large-Scale Pre-Ranking System: Entire- Chain Cross-Domain Models. In Proceedings of the 31st ACM I...

    2022

  17. [27]

    Jiaxi Tang and Ke Wang. 2018. Ranking Distillation: Learning Compact Ranking Models With High Performance for Recommender System. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Min- ing (London, United Kingdom) (KDD ’18). Associati...

    2018

  18. [28]

    Feng Wang, Jian Cheng, Weiyang Liu, and Haijun Liu. 2018. Additive Margin Softmax for Face Verification. IEEE Signal Processing Letters 25, 7 (2018), 926–930. https://doi.org/10.1109/LSP.2018.2822810

    2018

  19. [29]

    Jinpeng Wang, Jieming Zhu, and Xiuqiang He. 2021. Cross-Batch Negative Sampling for Training Two-Tower Recommenders. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval (Canada) (SIGIR ’21). Association for Computi...

    2021

  20. [30]

    X. Wang, H. Zhang, W. Huang, and M. R. Scott. 2020. Cross-Batch Memory for Embedding Learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) . IEEE Computer Society, Los Alamitos, CA, USA, 6387–6396. https://doi.org/10.1109/CVPR42600.2020.00642

    2020

  21. [31]

    Zhe Wang, Liqin Zhao, Biye Jiang, Guorui Zhou, Xiaoqiang Zhu, and Kun Gai

  22. [32]

    Fen Xia, Tie-Yan Liu, Jue Wang, Wensheng Zhang, and Hang Li. 2008. Listwise approach to learning to rank: theory and algorithm. In Proceedings of the 25th International Conference on Machine Learning (Helsinki, Finland) (ICML ’08) . Association for Computing Machinery, New Yor...

    2008

  23. [33]

    Chen Xu, Quan Li, Junfeng Ge, Jinyang Gao, Xiaoyong Yang, Changhua Pei, Fei Sun, Jian Wu, Hanxiao Sun, and Wenwu Ou. 2020. Privileged Features Distillation at Taobao Recommendations. arXiv:1907.05171 [cs.IR]

    2020 arXiv

  24. [34]

    Le Yan, Zhen Qin, Xuanhui Wang, Michael Bendersky, and Marc Najork. 2022. Scale Calibration of Deep Ranking Models. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining(Washington DC, USA)(KDD ’22). Association for Computing Machinery, New Yor...

    2022

  25. [35]

    Bianca Zadrozny. 2004. Learning and Evaluating Classifiers under Sample Selec- tion Bias. In Proceedings of the Twenty-First International Conference on Machine Learning (Banff, Alberta, Canada) (ICML ’04). Association for Computing Ma- chinery, New York, NY, USA, 114. https:/...

    2004

  26. [36]

    Yin Zhang, Ruoxi Wang, Derek Zhiyuan Cheng, Tiansheng Yao, Xinyang Yi, Lichan Hong, James Caverlee, and Ed H. Chi. 2023. Empowering Long-tail Item Recommendation through Cross Decoupling Network (CDN). In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and...

    2023

  27. [37]

    Zhixuan Zhang, Yuheng Huang, Dan Ou, Sen Li, Longbin Li, Qingwen Liu, and Xiaoyi Zeng. 2023. Rethinking the Role of Pre-ranking in Large-scale E-Commerce Searching System. arXiv:2305.13647 [cs.IR]

    2023 arXiv

  28. [38]

    Zhishan Zhao, Jingyue Gao, Yu Zhang, Shuguang Han, Siyuan Lou, Xiang-Rong Sheng, Zhe Wang, Han Zhu, Yuning Jiang, Jian Xu, and Bo Zheng. 2023. COPR: Consistency-Oriented Pre-Ranking for Online Advertising. In Proceedings of the 32nd ACM International Conference on Information ...

    2023

  29. [39]

    Guorui Zhou, Chengru Song, Xiaoqiang Zhu, Xiao Ma, Yanghui Yan, Xingya Dai, Han Zhu, Junqi Jin, Han Li, and Kun Gai. 2017. Deep Interest Network for Click-Through Rate Prediction. (2017)

    2017

  30. [2020]

    CoRR abs/2007.16122 (2020)

    COLD: Towards the Next Generation of Pre-Ranking System. CoRR abs/2007.16122 (2020). arXiv:2007.16122 https://arxiv.org/abs/2007.16122

    2020 arXiv

  31. [2023]

    arXiv preprint arXiv:2307.04322 (2023)

    Graph Contrastive Learning with Multi-Objective for Personalized Product Retrieval in Taobao Search. arXiv preprint arXiv:2307.04322 (2023)

    2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.