Pith. sign in

REVIEW 4 major objections 5 minor 52 references

Personalized Denoising Implicit Feedback for Robust Recommender System

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Per-user loss distributions separate real from noisy implicit feedback, and a resampling strategy built on them outperforms global-loss denoisers.

desk verdict Per-user loss separation is a genuine insight and PLD works on random-item noise; the main open question is whether it holds for systematic real-world noise. read the letter →

arxiv 2502.00348 v1 pith:K2SB3C7B submitted 2025-02-01 cs.IR

Kaike Zhang , Qi Cao , Yunfan Wu , Fei Sun , Huawei Shen , Xueqi Cheng This is my paper · ORCID
classification cs.IR
keywords robustrecommendersystemdenoisingrecommendationimplicitfeedbackpersonallossdistributionresamplingcollaborativefilteringoverlap
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

Recommenders trained on implicit feedback inherit clicks and views that do not always reflect true preference. Existing denoising methods assume noisy interactions have high loss in the global loss distribution and reweight or drop them, but this paper reports that the global distribution overlaps so heavily that such methods mislabel both clean and noisy interactions, and that the overlap worsens under pairwise BPR loss. The paper's central claim is that the confusion disappears when losses are viewed per user: for a fixed user, normal interactions consistently sit below noisy ones in that user's personal loss distribution. It therefore proposes PLD, a resampling strategy that draws each training positive from a small candidate pool of the user's own interactions with probability decreasing in loss, reducing the chance that noisy interactions are optimized. On Gowalla, Yelp2018, and MIND, plus a large MIND variant, with noise ratios from 0.1 to 0.5, PLD outperforms the evaluated reweighting and self-supervised denoisers, and a theoretical analysis predicts the same advantage.

What carries the argument

The central machinery is the user's personal loss distribution used as a sampling distribution. For each user $u$, PLD first constructs a candidate item pool $C_u^k$ by drawing $k$ items uniformly from that user's interacted items; it then computes each candidate's current training loss $l_{u,v}$ and resamples the positive item with probability $$P_{u,v}=\frac{\exp(-l_{u,v}/\tau)}{\sum_{j\in C_u^k}\exp(-l_{u,j}/\tau)}.$$ The temperature $\tau$ sharpens or flattens the softmax. Theorem 1 derives $\mathbb{E}[\Lambda_{\mathrm{normal}}-\Lambda_{\mathrm{noise}}]$, showing it exceeds the baseline $(n-m)/(n+m)$ for pool size $k>1$ and that smaller $\tau$ increases it. A variance correction for the softmax denominator enters through a fluctuation term that shrinks as the pool grows.

What would settle it

Run PLD on a dataset where noise is injected by replacing true positives with similar but wrong items; if many users then show overlapping personal loss quartiles for clean and noisy interactions, and PLD's Recall@20 no longer beats T-CE and R-CE on the same backbones, the paper's central claim would fail.

Watch

Extended reading notes

Core claim

The paper's discovery is that the overlap of clean and noisy implicit-feedback interactions in the global loss distribution is an artifact of pooling across users. Separating the losses per user, the two classes become distinguishable: the losses of normal interactions stay lower than those of noisy interactions for the same user, across noise ratios from 0.1 to 0.4. Because users differ in their overall loss scales, a global threshold cannot see this separation, which explains why high-loss reweighting and fixed drop rates misclassify a large fraction of both classes. PLD turns the per-user separation into a training procedure: uniformly sample a candidate pool from a user's interacted items, then resample one item with probability proportional to $\exp(-l_{u,v}/\tau)$. The paper proves a closed-form expression for $\mathbb{E}[\Lambda_{\mathrm{normal}}-\Lambda_{\mathrm{noise}}]$, the expected excess sampling probability of normal over noisy interactions, showing that PLD enlarges this gap beyond standard training and that lowering the temperature $\tau$ enlarges it further.

Load-bearing premise

The method assumes that real-world noise behaves like the random, user-independent injection used in the experiments, so that per-user clean and noisy losses remain separable; systematic noise such as misclicks on similar items would blur that separation.

Editorial extensions

If this is right

  • Denoisers that reweight or drop by global loss will keep misclassifying interactions under pairwise objectives, whereas PLD avoids a global threshold by construction.
  • PLD's per-user resampling removes the fixed drop-rate penalty for clean users: a user with no noise is never forced to discard interactions.
  • PLD adds negligible time overhead, roughly $O(kN)$ per epoch and no extra space, compared with sorting-based or bi-level denoisers.
  • The method transfers to pointwise BCE loss and can be stacked on contrastive-learning denoisers, so the per-user insight is not tied to a single objective or model family.

Reading between the lines

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

  • Beyond the paper, a testable extension is to inject systematic noise, such as misclicks on items similar to the user's true tastes, and measure whether per-user loss quartiles still separate; the paper only injects random items as noise, so this would probe whether the mechanism survives real-world error patterns.
  • The same per-user resampling idea could be applied to negative sampling and to user-specific loss variance, for example setting $\tau$ per user from an estimate of that user's loss spread rather than using one global value.
  • If PLD's mechanism is correct, denoising in collaborative filtering is less a data-cleaning step and more a per-user importance-sampling design, which suggests it should combine naturally with other samplers; the paper does not explore that composition.
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 / 5 minor

Summary. This paper studies denoising of implicit feedback in recommender systems. The authors observe that the overall loss distribution of normal and noisy interactions overlaps substantially, especially under BPR loss, and that this overlap is reduced when looking at per-user loss distributions. They propose PLD, which constructs a small candidate pool of a user's interacted items and then resamples one item per training step with probability proportional to exp(-loss/tau), thus down-weighting high-loss (likely noisy) interactions. A theoretical analysis (Theorem 1) is provided under Gaussian loss assumptions, and experiments on Gowalla, Yelp2018, MIND, and MIND-Large with MF and LightGCN backbones and both BPR and BCE losses report consistent improvements over several denoising baselines. The code is publicly available.

Significance. If the per-user separation between normal and noisy interaction losses holds for realistic noise, PLD is an attractive method: it is simple, adds negligible time and space overhead, works with both pointwise and pairwise losses, and improves over the compared baselines in the reported settings. The paper ships code and includes a formal-looking analysis of the sampling probabilities, which is a strength. The main risk is that the central separation premise is established only under randomly injected noise, and the theoretical result assumes the very ordering (mu1 < mu2) that the paper claims to discover. Missing statistical reporting (error bars, number of seeds) also makes the significance claims difficult to verify. These issues are addressable, however, and do not by themselves invalidate the empirical evidence.

major comments (4)
  1. [Section 4.1 (Figures 3-4, Tables 1-2) and Section 5.2 (Figure 5, Table 9)] The central empirical claim that normal and noisy interactions are separable in a user's personal loss distribution is established only for noise generated by randomly injected (user, item) pairs. Random items have essentially no affinity to the user, so the model cannot fit them and their loss is high almost by construction; the observed separation is therefore close to a tautology. Real implicit-feedback noise, such as misclicks or ambiguous feedback on similar items, can be semantically close to the user's preferences and may receive low loss, in which case the per-user ordering assumed by PLD may not hold. The paper should either add experiments with structured noise (for example, replacing items with same-category or nearest-embedding items, or using a real noisy-log dataset) and report the per-user loss overlap under that noise, or explicitly scope the method's guarantee to random-injection noise. This is the load-bearing evidence for the method, not a cosmetic addition.
  2. [Section 4.3 (Theorem 1, Eq. 2) and Appendix A.1] Theorem 1 assumes that normal interaction losses follow N(mu1, sigma^2) and noisy interaction losses follow N(mu2, sigma^2) with mu1 < mu2. The inequality mu1 < mu2 is exactly the separation that the paper claims to establish empirically; the theorem does not derive it from the model dynamics or from data properties. In addition, the proof relies on a Taylor expansion (Proposition 2) and on an unproven 'linear dependence' ansatz introducing a constant C in [beta, alpha], and the key inequality Gamma > chi / C^2 is asserted rather than proved. The validation in Figure 6 uses the same Gaussian assumptions and therefore checks the algebra of the approximation, not the validity of the assumptions. I recommend that the authors state Theorem 1 as a conditional result with explicit conditions, verify those conditions empirically during training (for example, by reporting estimated mu1, mu2, sigma per user), and either prove or numerically verify the fluctuation-term inequality for the range of n, m, k used in the experiments.
  3. [Section 5.1.4 (Implementation Details), Table 4, Figure 5] The paper reports t-test significance stars in Tables 4-6, 8, and 9, but gives no error bars, no standard deviations, and no number of seeds or runs. Many of the reported gains are only 1-3% in absolute terms, so without run-to-run variance information the reader cannot assess whether the improvements are within noise. Please report mean +/- std over at least five random seeds for all main tables and figures, and specify the exact t-test procedure (paired or unpaired, number of runs, and whether the test is over users or over runs).
  4. [Section 4.2 (Eq. 1 and Algorithm 1)] PLD uses the current model's own losses to resample the training data for the same model. The theoretical analysis treats the loss distributions as fixed Gaussians and does not analyze this feedback loop: if the model assigns low loss to certain interactions for idiosyncratic reasons, the resampling amplifies their presence, which can bias the training distribution and affect convergence. The paper should at least discuss this self-referential aspect and, ideally, provide a simple experiment that measures how the personal loss distributions evolve during PLD training and whether the separation persists throughout.
minor comments (5)
  1. [Section 5.3 heading and Figure 6] The heading 'Argumentation Study' appears to be a typo for 'Ablation Study', and the text 'randomly selecte 6 users' should read 'randomly selected 6 users'.
  2. [Introduction (Figure 1) and Section 4.1 (Tables 1-2)] The overlap percentages quoted in the Introduction (11.86%/98,596 normal and 11.52%/35,926 noisy for a 30% noise ratio) do not match the values in Table 1 (19.58%/162,763 and 18.16%/45,275) or Table 2 (4.28%/35,571 and 3.19%/7,969). Please reconcile these numbers and state which loss function and noise ratio each figure refers to.
  3. [Theorem 1, Eq. (2)] The notation C^2_k is ambiguous: it could be read as the square of the constant C times k, or as a binomial coefficient C(k,2), or as C^2 times k. Please define all symbols in the theorem statement.
  4. [Section 5.4 (Figure 8)] The hyperparameter analysis shows that performance is sensitive to tau, with tau = 0.05 recommended, but the theoretical discussion only says that decreasing tau enlarges the expected difference. A practical selection rule or a sensitivity table across datasets would help readers apply the method.
  5. [Section 4.1 (footnote)] The footnote says that similar experiments on different datasets and models yielded consistent results, but only the MIND/LightGCN configuration is shown. Please provide the corresponding overlap statistics for at least one additional dataset/backbone, or state where this material can be found.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found; the central evaluation is on held-out test data and the theoretical result is an explicit conditional statement, not a re-derivation of its own assumptions.

full rationale

The paper's derivation chain is not circular. The core empirical claim—that per-user loss distributions separate normal from noisy interactions—is established by experiments with injected random noise (Section 4.1), and the method itself resamples according to the model's own loss. This creates a training-time feedback loop, but the reported recommendation gains are measured on held-out test interactions (Tables 4-6, Figure 5), so the central evaluation is independent of the resampling mechanism. Theorem 1 is an explicitly conditional statement: it assumes μ1<μ2 and derives the sampling-probability advantage of PLD under that assumption; it does not purport to derive the separation from first principles. The assumption is load-bearing but not circular, because the separation is separately motivated empirically in Section 4.1. The use of random-item injection to define 'noisy' interactions is a potential threat to external validity for real-world systematic noise, but that is a correctness or robustness concern rather than a circularity. Self-citations (e.g., [41] and [42]-[44]) support background claims and dataset construction only; no load-bearing uniqueness theorem or fitted prediction is imported from the authors' prior work. Therefore, no specific circular step can be exhibited from the paper's own equations or citations.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The method rests on a few strong assumptions: Gaussian loss distributions in the theorem, random-item noise in experiments, and the persistence of per-user separation during training. The hyperparameters k and tau are tuned on validation sets.

free parameters (2)
  • Candidate pool size k = selected from {2,3,5,10,20} per dataset
    Controls trade-off between sampling variance and stability; tuned on validation set.
  • Temperature coefficient tau = selected from {0.01,0.05,0.1,0.2,0.3,0.4,0.5}
    Controls sharpness of resampling distribution; theoretical analysis suggests smaller tau increases separation but experiments show optimum around 0.05.
assumptions (3)
  • ad hoc to paper Loss of normal interactions follows N(mu1, sigma^2) and loss of noisy interactions follows N(mu2, sigma^2) with mu1 < mu2 and mu1,mu2 > sigma.
    Used in Appendix A.1 to derive Theorem 1; not validated on actual loss distributions.
  • domain assumption Noisy interactions are random items added to the user's interaction set.
    Used in all experiments; may not represent real-world noise like misclicks or ambiguous feedback.
  • domain assumption The per-user distinction between normal and noisy losses persists during training and after resampling.
    The method relies on this to keep filtering noise; not formally proven and could be weakened by the feedback loop.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Personalized Denoising Implicit Feedback for Robust Recommender System." pith.science (2026). https://pith.science/paper/K2SB3C7B

@misc{pith2026250200348,
  author       = {Pith},
  title        = {Pith review of: Personalized Denoising Implicit Feedback for Robust Recommender System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K2SB3C7B}},
  note         = {Machine review of arXiv:2502.00348}
}
read the original abstract

While implicit feedback is foundational to modern recommender systems, factors such as human error, uncertainty, and ambiguity in user behavior inevitably introduce significant noise into this feedback, adversely affecting the accuracy and robustness of recommendations. To address this issue, existing methods typically aim to reduce the training weight of noisy feedback or discard it entirely, based on the observation that noisy interactions often exhibit higher losses in the overall loss distribution. However, we identify two key issues: (1) there is a significant overlap between normal and noisy interactions in the overall loss distribution, and (2) this overlap becomes even more pronounced when transitioning from pointwise loss functions (e.g., BCE loss) to pairwise loss functions (e.g., BPR loss). This overlap leads traditional methods to misclassify noisy interactions as normal, and vice versa. To tackle these challenges, we further investigate the loss overlap and find that for a given user, there is a clear distinction between normal and noisy interactions in the user's personal loss distribution. Based on this insight, we propose a resampling strategy to Denoise using the user's Personal Loss distribution, named PLD, which reduces the probability of noisy interactions being optimized. Specifically, during each optimization iteration, we create a candidate item pool for each user and resample the items from this pool based on the user's personal loss distribution, prioritizing normal interactions. Additionally, we conduct a theoretical analysis to validate PLD's effectiveness and suggest ways to further enhance its performance. Extensive experiments conducted on three datasets with varying noise ratios demonstrate PLD's efficacy and robustness.

Figures

Figures reproduced from arXiv: 2502.00348 by the authors.

Figure 1
Figure 1. Probability distribution of losses. The overlap re [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 4
Figure 4. Difference between normal and noisy interactions [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 3
Figure 3. Personal loss distribution for five users. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Recommendation performance of different denoising methods across various noise ratios. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 7
Figure 7. Figure 7: Number of normal interactions and noisy interac [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Left: Analysis of hyper-parameter 𝑘; Right: Analysis of hyper-parameter 𝜏 [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Training time per epoch (in seconds) with batch [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 41 canonical work pages

  1. [1]

    Kai Lai Chung. 2000. A Course in Probability Theory . Elsevier

    work page 2000

  2. [2]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep Neural Networks for Youtube Recommendations. In Proceedings of the 10th ACM Conference on Recommender Systems. 191–198

    work page 2016

  3. [3]

    Ziwei Fan, Ke Xu, Zhang Dong, Hao Peng, Jiawei Zhang, and Philip S Yu. 2023. Graph Collaborative Signals Denoising and Augmentation for Recommendation. In Proceedings of the 46th international ACM SIGIR Conference on Research and Development in Information Retrieval . 2037–2041

    work page 2023

  4. [4]

    Zeno Gantner, Lucas Drumond, Christoph Freudenthaler, and Lars Schmidt- Thieme. 2012. Personalized Ranking for Non-uniformly Sampled Items. In Pro- ceedings of KDD Cup 2011 . PMLR, 231–247

    work page 2012

  5. [5]

    Yunjun Gao, Yuntao Du, Yujia Hu, Lu Chen, Xinjun Zhu, Ziquan Fang, and Baihua Zheng. 2022. Self-Guided Learning to Denoise for Robust Recommendation. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval . ACM, 1412–1422

    work page 2022

  6. [6]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, YongDong Zhang, and Meng Wang. 2020. LightGCN - Simplifying and Powering Graph Convolution Net- work for Recommendation. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval . 639–648

    work page 2020

  7. [7]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural Collaborative Filtering. InProceedings of the 26th International Conference on World Wide Web. 173–182

    2017

  8. [8]

    Zhuangzhuang He, Yifan Wang, Yonghui Yang, Peijie Sun, Le Wu, Haoyue Bai, Jinqi Gong, Richang Hong, and Min Zhang. 2024. Double Correction Frame- work for Denoising Recommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , Vol. 33. ACM, 1062–1072

    work page 2024

Show all 52 references
  1. [9]

    Jonathan L Herlocker, Joseph A Konstan, Loren G Terveen, and John T Riedl. 2004. Evaluating Collaborative Filtering Recommender Systems. ACM Transactions on Information Systems (TOIS) 22, 1 (2004), 5–53

    2004

  2. [10]

    Yifan Hu, Yehuda Koren, and Chris Volinsky. 2008. Collaborative Filtering for Implicit Feedback Datasets. In 2008 8th IEEE International Conference on Data Mining. IEEE, 263–272

    2008

  3. [11]

    Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix Factorization Tech- niques for Recommender Systems. Computer 42, 8 (2009), 30–37

    2009

  4. [12]

    Yehuda Koren, Steffen Rendle, and Robert Bell. 2021. Advances in collaborative filtering. Recommender systems handbook (2021), 91–142

    2021

  5. [13]

    Erich L Lehmann and George Casella. 2006. Theory of Point Estimation . Springer Science & Business Media

    2006

  6. [14]

    Dawen Liang, Laurent Charlin, James McInerney, and David M Blei. 2016. Mod- eling User Exposure in Recommendation. In Proceedings of the 25th International Conference on World Wide Web. 951–961

    2016

  7. [15]

    Weilin Lin, Xiangyu Zhao, Yejing Wang, Yuanshao Zhu, and Wanyu Wang. 2023. Autodenoise: Automatic data instance denoising for recommendations. In Pro- ceedings of the ACM Web Conference 2023 . 1003–1011

    2023

  8. [16]

    Wenze Ma, Yuexian Wang, Yanmin Zhu, Zhaobo Wang, Mengyuan Jing, Xuhao Zhao, Jiadi Yu, and Feilong Tang. 2024. MADM: A Model-agnostic Denoising Module for Graph-based Social Recommendation. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining . 501–509

    2024

  9. [17]

    Yuhan Quan, Jingtao Ding, Chen Gao, Lingling Yi, Depeng Jin, and Yong Li. 2023. Robust Preference-guided Denoising for Graph Based Social Recommendation. In Proceedings of the 32nd International Conference on World Wide Web . 1097–1108

    2023

  10. [18]

    Xubin Ren, Lianghao Xia, Jiashu Zhao, Dawei Yin, and Chao Huang. 2023. Disen- tangled contrastive collaborative filtering. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval . 1137–1146

    2023

  11. [19]

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

  12. [20]

    John A Rice and John A Rice. 2007. Mathematical Statistics and Data Analysis . Vol. 371. Thomson/Brooks/Cole Belmont, CA

    2007

  13. [21]

    Yuta Saito, Suguru Yaginuma, Yuta Nishino, Hayato Sakata, and Kazuhide Nakata

  14. [22]

    J Ben Schafer, Dan Frankowski, Jon Herlocker, and Shilad Sen. 2007. Collaborative Filtering Recommender Systems. In The Adaptive Web: Methods and Strategies of Web Personalization. Springer, 291–324

    2007

  15. [23]

    Changxin Tian, Yuexiang Xie, Yaliang Li, Nan Yang, and Wayne Xin Zhao. 2022. Learning to Denoise Unreliable Interactions for Graph Collaborative Filtering. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval . ACM, 122–132

    2022

  16. [24]

    Raciel Yera Toledo, Jorge Castro, and Luis Martínez-López. 2016. A fuzzy Model for Managing Natural Noise in Recommender Systems. Applied Soft Computing 40 (2016), 187–198

    2016

  17. [25]

    Hao Wang, Naiyan Wang, and Dit-Yan Yeung. 2015. Collaborative Deep Learning for Recommender Systems. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining . 1235–1244

    2015

  18. [26]

    Pengfei Wang, Chenliang Li, Lixin Zou, Zhichao Feng, Kaiyuan Li, Xiaochen Li, Xialong Liu, and Shangguang Wang. 2023. Tutorial: Data Denoising Metrics in Recommender Systems. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management . 5224–5227

    2023

  19. [27]

    Wenjie Wang, Fuli Feng, Xiangnan He, Liqiang Nie, and Tat-Seng Chua. 2021. Denoising Implicit Feedback for Recommendation. In Proceedings of the 14th ACM International Conference on Web Search and Data Mining . ACM, 373–381

    2021

  20. [28]

    Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural Graph Collaborative Filtering. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval . 165–174

    2019

  21. [29]

    Xiang Wang, Hongye Jin, An Zhang, Xiangnan He, Tong Xu, and Tat-Seng Chua

  22. [30]

    Yu Wang, Xin Xin, Zaiqiao Meng, Joemon M Jose, Fuli Feng, and Xiangnan He

  23. [31]

    Zongwei Wang, Min Gao, Wentao Li, Junliang Yu, Linxin Guo, and Hongzhi Yin. 2023. Efficient Bi-Level Optimization for Recommendation Denoising. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. ACM, 2502–2511

    2023

  24. [32]

    In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval

    Disentangled Graph Collaborative Filtering. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. 1001–1010

  25. [33]

    Jiancan Wu, Xiang Wang, Fuli Feng, Xiangnan He, Liang Chen, Jianxun Lian, and Xing Xie. 2021. Self-supervised Graph Learning for Recommendation. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval . 726–735

    2021

  26. [34]

    Shiwen Wu, Fei Sun, Wentao Zhang, Xu Xie, and Bin Cui. 2022. Graph Neural Networks in Recommender Systems: A Survey. Comput. Surveys 55, 5 (2022), 1–37

    2022

  27. [35]

    Yao Wu, Christopher DuBois, Alice X Zheng, and Martin Ester. 2016. Collabora- tive Denoising Auto-encoders for Top-n Recommender Systems. InProceedings of the 9th ACM International Conference on Web Search and Data Mining . 153–162

    2016

  28. [36]

    Fangzhao Wu, Ying Qiao, Jiun-Hung Chen, Chuhan Wu, Tao Qi, Jianxun Lian, Danyang Liu, Xing Xie, Jianfeng Gao, Winnie Wu, et al. 2020. Mind: A Large-scale Dataset for News Recommendation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics...

    2020

  29. [37]

    Yuhao Yang, Chao Huang, Lianghao Xia, and Chenliang Li. 2022. Knowledge Graph Contrastive Learning for Recommendation. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, 1434–1443

    2022

  30. [38]

    Haibo Ye, Xinjie Li, Yuan Yao, and Hanghang Tong. 2023. Towards Robust Neural Graph Collaborative Filtering via Structure Denoising and Embedding Perturbation. ACM Transactions on Information Systems 41, 3 (2023), 1–28

    2023

  31. [39]

    Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L Hamilton, and Jure Leskovec. 2018. Graph Convolutional Neural Networks for Web-scale Recommender Systems. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining . 974–983

    2018

  32. [40]

    Lianghao Xia, Chao Huang, Yong Xu, Jiashu Zhao, Dawei Yin, and Jimmy Huang

  33. [41]

    In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Hypergraph Contrastive Collaborative Filtering. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. 70–79

  34. [42]

    Kaike Zhang, Qi Cao, Yunfan Wu, Fei Sun, Huawei Shen, and Xueqi Cheng. 2024. Improving the shortest plank: Vulnerability-aware adversarial training for robust recommender system. In Proceedings of the 18th ACM Conference on Recommender Systems. 680–689

    2024

  35. [43]

    Kaike Zhang, Qi Cao, Yunfan Wu, Fei Sun, Huawei Shen, and Xueqi Cheng. 2024. LoRec: Combating Poisons with Large Language Model for Robust Sequential Recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrie...

    2024

  36. [44]

    Kaike Zhang, Qi Cao, Yunfan Wu, Fei Sun, Huawei Shen, and Xueqi Cheng. 2024. Understanding and Improving Adversarial Collaborative Filtering for Robust Recommendation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  37. [45]

    Wenhui Yu and Zheng Qin. 2020. Sampler Design for Implicit Feedback Data by Noisy-label Robust Learning. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval . 861–870

    2020

  38. [46]

    Kaike Zhang, Qi Cao, Fei Sun, Yunfan Wu, Shuchang Tao, Huawei Shen, and Xueqi Cheng. 2023. Robust Recommender System: A Survey and Future Directions. arXiv preprint arXiv:2309.02057 (2023)

    2023 arXiv

  39. [50]

    Jujia Zhao, Wang Wenjie, Yiyan Xu, Teng Sun, Fuli Feng, and Tat-Seng Chua. 2024. Denoising Diffusion Recommender Model. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval . 1370–1379

    2024

  40. [51]

    Xinjun Zhu, Yuntao Du, Yuren Mao, Lu Chen, Yujia Hu, and Yunjun Gao. 2023. Knowledge-refined Denoising Network for Robust Recommendation. In Proceed- ings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 362–371. WWW ’25, Apr...

    2023

  41. [52]

    Cov 𝑁∑︁ 𝑖=1 exp(−𝑥𝑖)| 𝑁, 1 𝑆𝑥+𝑆𝑦 !# + Cov E

    represent the loss of a normal sample𝑖, and𝑦𝑗∼N( 𝜇2,𝜎 2) represent the loss of a noisy sample 𝑗. According to Equation 1, the probability of selecting sample𝑖 is: 𝑃𝑖 = exp(−𝑥𝑖) Í𝑁 𝑖=1 exp(−𝑥𝑖)+ Í𝑀 𝑗=1 exp(−𝑦𝑗) . Define: 𝑆𝑥 = 𝑁∑︁ 𝑖=1 exp(−𝑥𝑖), 𝑆 𝑦 = 𝑀∑︁ 𝑗=1 exp(−𝑦𝑗). Personaliz...

    2025

  42. [2009]

    InProceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence

    BPR: Bayesian Personalized Ranking from Implicit Feedback. InProceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence . 452–461

  43. [2020]

    In Proceedings of the 13th International Conference on Web Search and Data Mining

    Unbiased recommender learning from missing-not-at-random implicit feedback. In Proceedings of the 13th International Conference on Web Search and Data Mining. 501–509

  44. [2022]

    In Proceedings of the ACM Web Conference 2022

    Learning Robust Recommenders through Cross-Model Agreement. In Proceedings of the ACM Web Conference 2022 . ACM, 2015–2025

    2022

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.