Pith. sign in

REVIEW 4 major objections 5 minor 43 references

Unlocking the Hidden Treasures: Enhancing Recommendations with Unlabeled Data

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · 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 collaborative-filtering models with a three-way partial ranking — positive above neutral above negative — consistently improves top-K recommendation across a range of base models.

desk verdict A solid empirical recipe with a genuinely new three-class ranking idea, but the theory is shaky and the deployed loss is not the set-level ranking the paper claims. read the letter →

arxiv 2412.18170 v1 pith:2AGMEP3X submitted 2024-12-24 cs.IR

Yuhan Zhao , Rui Chen , Qilong Han , Hongtao Song , Li Chen This is my paper · ORCID
classification cs.IR
keywords collaborativefilteringpositive-neutral-negativelearningpartialrankingunlabeleddataone-wayAUCsemi-supervisedtop-Krecommendationlossfunction
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 sets out to change how collaborative filtering treats the vast pool of unlabeled interactions. Instead of sampling some of them as negatives and discarding the rest, it introduces a third 'neutral' class for items that are neither clearly liked nor clearly disliked, and trains models to satisfy the partial ranking positive > neutral > negative. The central claim is that this three-way ranking makes better use of unlabeled data and consistently improves top-K recommendation across standard base models. The authors also derive a link between the PNN objective and one-way partial AUC, which gives a theoretical reason to expect gains on ranking metrics. If the claim holds, a simple matrix-factorization model with PNN can reach the recommendation quality of much heavier graph-neural-network models.

What carries the argument

The machinery is the PNN loss $L_{PNN} = \alpha L_{constrain} + \beta L_{uniform} + L_{rank}$. First, a BPR warm-start phase with batch-wise dynamic negative sampling learns enough preference signal to classify unlabeled items; a user-aware attention model sets $\lambda$ to shift weight from BPR to PNN as classification confidence grows. Then the ranking term uses two-step centroid ranking: each class is summarized by its mean embedding, and BPR-style losses enforce centroids in the order positive > neutral > negative, while two clamp embeddings squeeze the neutral class so individual items separate from the other classes, and a uniformity loss spreads unlabeled items apart to make classification easier. A mini-batch approximation treats other users' items in the batch as the neutral and negative sets for each user.

What would settle it

Compare PNN against a control that randomly splits unlabeled items into neutral and negative classes with the same sizes; if the control matches PNN's top-K scores, the classification step is not the source of the gain. Separately, run a full-batch variant on a small dataset where every user's unlabeled items are visible; a large drop in NDCG@K relative to the mini-batch version would show that the Eq. 18 approximation does not preserve the claimed set-level ranking.

Watch

Extended reading notes

Core claim

The paper's central claim is that the standard practice of treating all unlabeled interactions as negatives throws away exactly the ambiguous items that would sharpen user-preference learning. PNN inserts a neutral class between positive and negative and optimizes the set-level partial ranking $\Sigma_{pos} > \Sigma_{neu} > \Sigma_{neg}$. The authors derive that this objective tracks one-way partial AUC, so optimizing it targets the false-positive-rate regime that matters for top-K metrics. The reported experiments show consistent gains when PNN is swapped into BPR-MF, LightGCN, NGCF, and SGL, and the headline comparison is that BPR-MF plus PNN reaches the level of sophisticated graph-based models.

Load-bearing premise

The load-bearing premise is that training on small batches of other users' items faithfully stands in for ranking a user against all of her unlabeled items; if that substitution fails, the method's gains may come from its auxiliary regularization terms rather than from the neutral class itself.

Editorial extensions

If this is right

  • Across the four base models and four datasets, PNN improves nearly all reported metrics; the average improvements range from roughly 3% to 30%, with one combination (SGL on Foursquare) showing no net gain.
  • BPR-MF with PNN reaches comparable performance to graph-based models like LightGCN and NGCF without PNN, so the paradigm can substitute for architectural complexity.
  • PNN outperforms comparison methods such as SimpleX, UIB, SRNS, MixGCF, and ANS across all four datasets on the reported metrics.
  • When 10% of interactions are deliberately corrupted to simulate false negatives, PNN retains more performance than the comparison methods, consistent with the neutral class absorbing noisy items.
  • Ablations show that removing the ranking term or the semi-supervised classification step causes the largest performance drop.

Reading between the lines

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

  • The batch-level approximation in Eq. 18 means the deployed loss never sees a user's full item sets; an implication the paper leaves open is that the gains may partly come from the uniformity and clamp regularizers rather than from a true set-level ranking.
  • PNN could generalize to more than one intermediate class, for example confidence-weighted neutral items, since the centroid ranking scheme does not depend on exactly three classes.
  • The user-aware attention weight $\lambda$ is a plausible but untested proxy for classification readiness; a simpler schedule based on validation performance might work as well and would isolate the benefit of the attention model.
  • A direct test of the neutral-class hypothesis would use explicit 'maybe' feedback, such as middling ratings or later re-purchases, as ground truth for the neutral class; the paper's synthetic false-negative experiment is a partial version of that test.
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. The paper introduces the positive-neutral-negative (PNN) learning paradigm for collaborative filtering, in which unlabeled items are stratified into positive, neutral, and negative classes and the model is trained on the set-level partial ranking Σ_pos > Σ_neu > Σ_neg. The authors provide a theoretical argument connecting PNN to one-way partial AUC (OPAUC), a concrete implementation based on semi-supervised classification with a user-aware attention mechanism, a two-step centroid ranking loss with a clamp mechanism, and a mini-batch relaxation. Experiments on four real-world datasets with four base CF models and comparisons to several state-of-the-art negative-sampling methods report consistent improvements, with code publicly available.

Significance. If the proposed mechanism is validated, PNN offers a model-agnostic way to exploit the full unlabeled data rather than discarding the portion not selected as negative samples, and the empirical results are extensive: four datasets, four base models, standard top-K metrics, ablations, hyperparameter studies, and efficiency comparisons. The public code and use of RecBole are strengths for reproducibility. However, the theoretical derivation in Section 2 and the connection between the deployed loss and the claimed set-level ranking are not established, so the significance is conditional on fixing these gaps.

major comments (4)
  1. [§2, Eqs. (5)–(6)] The OPAUC derivation is not valid as written. In Eq. (5), the index j in the indicator I(r_{ui} > r_{uj}) is undefined, and the subsequent claim that 'it is evident that r_{u i_n} > r_{u j}' is not generally true when j ranges over unlabeled data, which may contain positive or neutral items. The transition from Eq. (5) to Eq. (6) also replaces a pairwise comparison with a triple-wise comparison without justification. Since this derivation is the paper's theoretical validation, it needs to be corrected or the claim should be weakened.
  2. [§3.3, Eq. (18)] The mini-batch relaxation is asserted rather than derived. The equality between the product over positive items of per-item probabilities against batch-level neutral and negative sets and the global set-level ranking probability requires independence and coverage conditions that are not stated or proved. Because the deployed loss uses this relaxation, the reported performance gains cannot be attributed to the claimed triple-wise objective without additional evidence that the approximation is accurate.
  3. [§3.2.2, Eqs. (14) and (16)] The text claims that the two-step centroid ranking and clamp mechanism 'guarantees that every item in a class ranks higher or lower than any item in another class,' but Eq. (14) only supervises centroid scores and Eq. (16) only minimizes the distance between two auxiliary clamp embeddings. No argument or measurement shows that these operations enforce the per-item set-level inequalities, which are strictly stronger than centroid separation.
  4. [Table 3, Yelp row for MixGCF] The reported MRR@10 for MixGCF on Yelp is 0.4680, identical to the ML-1M MixGCF MRR@10 and roughly six times larger than all other Yelp MRR values. As printed, this contradicts the claim in §4.2 that 'PNN consistently delivers the best performance across all datasets and all metrics,' because PNN's Yelp MRR@10 is 0.0739. Please verify the baseline tables and correct any transcription errors.
minor comments (5)
  1. [Abstract and §1] The phrase 'graph neutral networks' should be 'graph neural networks'.
  2. [§2, Eq. (4)] The notation S_{I-_u}[1, |I-_u|·δ] is not defined for non-integer |I-_u|·δ, and the subsequent use of S_{I_u^nu} in Eq. (5) is unclear.
  3. [§3.1, Eq. (10)] The user-aware attention model is motivated by the intuition that high attention weights indicate correct identification of positive items, but the mechanism by which attention scores measure classification performance is not made explicit; a brief formal justification would improve clarity.
  4. [Table 2, Foursquare SGL] The Foursquare SGL results show decreases in Hit@20 and NDCG@20 when PNN is added, and the improvement is reported as '-'. The text in §4.2 says improvements occur 'in almost all cases,' but the failure case should be acknowledged and discussed explicitly.
  5. [§4.1.3] The statement that findings are based on five independent runs and p-values are calculated against the best-performing baseline is useful, but the specific p-values are not reported; providing them or a significance table would strengthen the claims.

Circularity Check

1 steps flagged · score 4.0 of 10

The OPAUC validation is a notational restatement of the PNN objective; the empirical claim is independent.

  1. self definitional [Section 2, Equations (5)–(6), leading to Equation (2)]
    "Then, for these samples, it is evident that r_ui_n > r_uj. Consequently, we can rewrite the equation as OPAUC(δ) =|K| Σ_{i∈I+u} Σ_{i_n∈Iunu} Σ_{j∈Iunu} I(r_ui > r_ui_n > r_uj) I(i_n∈S_I-_u) ... Obviously, by expressing Eq. 6 in set notation and introducing symbols for positive, neutral, and negative classes within PNN, we arrive at Eq. 2."

    Equation (6) is the triple-wise ranking r_pos > r_neutral > r_neg (with the 'neutral' class identified as top-ranked unlabeled items). Equation (2) is exactly that set-level ranking rewritten in PNN notation. The derivation does not import independent content from OPAUC: it inserts the PNN ordering into the OPAUC expression via the assertion 'it is evident that r_ui_n > r_uj' and then reads Eq. (2) back out via 'obviously'. Thus the claimed 'theoretical connection to OPAUC' is a restatement of the PNN objective by construction. Moreover, the implemented neutral/negative split (greedy lowest-score negatives in Section 3.2.1) is not shown to coincide with the top-ranked-negative set that OPAUC requires, so even the restatement's applicability to the deployed loss is unsupported.

full rationale

The only circularity identified is localized in Section 2: the 'theoretical analysis' connecting PNN to OPAUC reduces to re-expressing the triple-wise ranking that defines PNN, so the validation is partly self-definitional. The paper's central empirical claim—consistent, significant top-K gains across four base models and four datasets (Tables 2 and 3)—is evaluated against external baselines and does not rely on the theoretical section, so it retains independent content. The mini-batch approximation of Eq. (18) and the clamp mechanism's claimed guarantee of set-level separation are unproven and could disconnect the deployed loss from the stated objective, but these are correctness risks, not circularity. No load-bearing self-citations or fitted-parameter-renamed-as-prediction issues were found. Overall circularity is moderate and confined to the theoretical motivation.

Assumptions & free parameters 3 free parameters · 6 assumptions · 2 invented entities

The method relies on a set of domain assumptions about the structure of unlabeled data, plus ad hoc components (clamp, uniformity) whose exact form is not derived from first principles. The free hyperparameters α and β are tuned per dataset and not reported.

free parameters (3)
  • α (weight of U_constrain) = not reported per dataset (tuned in [0,1])
    Eq. 17; balances the clamp constraint. Tuned on validation for each dataset, but exact values are omitted.
  • β (weight of U_uniform) = not reported per dataset (tuned in [0,1])
    Eq. 17; balances the uniformity term. Tuned on validation.
  • Clamp noise magnitude = 0.1 (upper bound of uniform distribution)
    Eq. 15; Δ ~ U(0,0.1); chosen by hand, no sensitivity analysis.
assumptions (6)
  • domain assumption Unlabeled data can be partitioned into neutral and negative classes, and the triple-wise ordering Σ_pos > Σ_neu > Σ_neg reflects user preferences.
    Section 2, Eq. 2; the paper's core modeling premise.
  • domain assumption The lowest-similarity unlabeled items are reliably true negatives.
    Section 3.2.1, Eq. 12; the greedy selection rule.
  • domain assumption The centroid embedding of a class preserves the set-level ranking relationships between classes.
    Section 3.2.2, after Eq. 13; used to turn set-level ranking into item-level BPR-like losses.
  • ad hoc to paper The clamp mechanism with uniform noise on the direction of the positive/negative centroid separates classes without harmful inductive bias.
    Section 3.2.2, Eqs. 15-16; no prior evidence that this exactly creates correct margins.
  • ad hoc to paper The product over positive items in mini-batches approximates the full set-level ranking.
    Section 3.3, Eq. 18; unproven approximation that enables mini-batch training.
  • domain assumption BPR loss can warm-start the model to acquire enough knowledge to classify neutral vs negative items.
    Section 3.1; relies on BPR's pairwise ranking to produce discriminative scores.
invented entities (2)
  • Neutral class (Σ_neu)
    purpose: Latent class for unlabeled items that are neither clearly positive nor negative, enabling triple-wise ranking.
    A modeling construct; no external signal or falsifiable prediction identifies neutral items.
  • Clamp embeddings (e_clamp)
    purpose: Artificial boundary embeddings used to create adaptive margins between positive-neutral and neutral-negative classes.
    Introduced in Section 3.2.2; internal to the loss, no independent validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unlocking the Hidden Treasures: Enhancing Recommendations with Unlabeled Data." pith.science (2026). https://pith.science/paper/2AGMEP3X

@misc{pith2026241218170,
  author       = {Pith},
  title        = {Pith review of: Unlocking the Hidden Treasures: Enhancing Recommendations with Unlabeled Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2AGMEP3X}},
  note         = {Machine review of arXiv:2412.18170}
}
read the original abstract

Collaborative filtering (CF) stands as a cornerstone in recommender systems, yet effectively leveraging the massive unlabeled data presents a significant challenge. Current research focuses on addressing the challenge of unlabeled data by extracting a subset that closely approximates negative samples. Regrettably, the remaining data are overlooked, failing to fully integrate this valuable information into the construction of user preferences. To address this gap, we introduce a novel positive-neutral-negative (PNN) learning paradigm. PNN introduces a neutral class, encompassing intricate items that are challenging to categorize directly as positive or negative samples. By training a model based on this triple-wise partial ranking, PNN offers a promising solution to learning complex user preferences. Through theoretical analysis, we connect PNN to one-way partial AUC (OPAUC) to validate its efficacy. Implementing the PNN paradigm is, however, technically challenging because: (1) it is difficult to classify unlabeled data into neutral or negative in the absence of supervised signals; (2) there does not exist any loss function that can handle set-level triple-wise ranking relationships. To address these challenges, we propose a semi-supervised learning method coupled with a user-aware attention model for knowledge acquisition and classification refinement. Additionally, a novel loss function with a two-step centroid ranking approach enables handling set-level rankings. Extensive experiments on four real-world datasets demonstrate that, when combined with PNN, a wide range of representative CF models can consistently and significantly boost their performance. Even with a simple matrix factorization, PNN can achieve comparable performance to sophisticated graph neutral networks.

Figures

Figures reproduced from arXiv: 2412.18170 by the authors.

Figure 1
Figure 1. The overall workflow of our PNN learning paradigm. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The overall process of the clamp mechanism. pos, [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The results of the study on neutral class. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Impact of hyperparameters on PNN. We conduct experiments on Foursquare and Yelp to study the impact of different values of 𝛼 and 𝛽 and present the results in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 38 canonical work pages

  1. [1]

    Walter E Block, II Barnett, and Others. 2012. Transitivity and The Money Pump. Quarterly Journal of Austrian Economics 15, 2 (2012)

    work page 2012

  2. [2]

    Chong Chen, Min Zhang, Chenyang Wang, Weizhi Ma, Minming Li, Yiqun Liu, and Shaoping Ma. 2019. An Efficient Adaptive Transfer Neural Network for Social-Aware Recommendation. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval . 225–234

    work page 2019

  3. [3]

    Chong Chen, Min Zhang, Yongfeng Zhang, Yiqun Liu, and Shaoping Ma. 2020. Efficient Neural Matrix Factorization without Sampling for Recommendation. ACM Transactions on Information Systems (TOIS) 38, 2 (2020), 1–28

    work page 2020

  4. [4]

    Jin Chen, Defu Lian, Binbin Jin, Kai Zheng, and Enhong Chen. 2022. Learning Recommenders for Implicit Feedback with Importance Resampling. InProceedings of the ACM Web Conference. 1997–2005

    work page 2022

  5. [5]

    Jingtao Ding, Yuhan Quan, Xiangnan He, Yong Li, and Depeng Jin. 2019. Rein- forced Negative Sampling for Recommendation with Exposure Data.. In Interna- tional Joint Conference on Artificial Intelligence . 2230–2236

    work page 2019

  6. [6]

    Jingtao Ding, Yuhan Quan, Quanming Yao, Yong Li, and Depeng Jin. 2020. Simplify and Robustify Negative Sampling for Implicit Collaborative Filtering. Advances in Neural Information Processing Systems 33 (2020), 1094–1105

    work page 2020

  7. [7]

    Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. 2014. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572 (2014)

    arXiv 2014

  8. [8]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang

Show all 43 references
  1. [9]

    Xiangnan He, Zhankui He, Xiaoyu Du, and Tat-Seng Chua. 2018. Adversarial Personalized Ranking for Recommendation. In The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval . 355–364

    2018

  2. [10]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua

  3. [11]

    Tinglin Huang, Yuxiao Dong, Ming Ding, Zhen Yang, Wenzheng Feng, Xinyu Wang, and Jie Tang. 2021. MixGCF: An Improved Training Method for Graph Neural Network-based Recommender Systems. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 665–674

    2021

  4. [12]

    Yangqin Jiang, Chao Huang, and Lianghao Xia. 2023. Adaptive Graph Contrastive Learning for Recommendation. arXiv preprint arXiv:2305.10837 (2023)

    2023 arXiv

  5. [13]

    Riwei Lai, Li Chen, Yuhan Zhao, Rui Chen, and Qilong Han. 2023. Disentangled Negative Sampling for Collaborative Filtering. In Proceedings of the 16th ACM International Conference on Web Search and Data Mining . 96–104

    2023

  6. [14]

    Guanyu Lin, Chen Gao, Yu Zheng, Yinfeng Li, Jianxin Chang, Yanan Niu, Yang Song, Kun Gai, Zhiheng Li, Depeng Jin, et al . 2024. Inverse Learning with Extremely Sparse Feedback for Recommendation. In Proceedings of the 17th ACM International Conference on Web Search and Data Mi...

    2024

  7. [15]

    Kelong Mao, Jieming Zhu, Jinpeng Wang, Quanyu Dai, Zhenhua Dong, Xi Xiao, and Xiuqiang He. 2021. SimpleX: A Simple and Strong Baseline for Collaborative Filtering. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management. 1243–1252

    2021

  8. [16]

    Lucas Maystre and Matthias Grossglauser. 2017. Just Sort It! A Simple and Effective Approach to Active Preference Learning. In Proceedings of the 34th International Conference on Machine Learning , Vol. 70. PMLR, 2344–2353

    2017

  9. [17]

    Steffen Rendle and Christoph Freudenthaler. 2014. Improving Pairwise Learning for Item Recommendation from Implicit Feedback. In Proceedings of the 7th ACM International Conference on Web Search and Data Mining . 273–282. RecSys ’24, October 14–18, 2024, Bari, Italy Yuhan Zhao...

    2014

  10. [18]

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

  11. [19]

    Mamshad Nayeem Rizve, Kevin Duarte, Yogesh S Rawat, and Mubarak Shah. 2021. In Defense of Pseudo-Labeling: An Uncertainty-aware Pseudo-label Selection Famework for Semi-Supervised Learning. arXiv preprint arXiv:2101.06329 (2021)

    2021 arXiv

  12. [20]

    Wentao Shi, Jiawei Chen, Fuli Feng, Jizhi Zhang, Junkang Wu, Chongming Gao, and Xiangnan He. 2023. On the Theories behind Hard Negative Sampling for Recommendation. In Proceedings of the ACM Web Conference . 812–822

    2023

  13. [21]

    Trevor Standley, Amir Zamir, Dawn Chen, Leonidas Guibas, Jitendra Malik, and Silvio Savarese. 2020. Which Tasks Should be Learned Together in Multi-Task Learning?. In International Conference on Machine Learning . 9120–9132

    2020

  14. [22]

    Xiaoyuan Su and Taghi M Khoshgoftaar. 2009. A Survey of Collaborative Filtering Techniques. Advances in Artificial Intelligence 2009 (2009)

    2009

  15. [23]

    Jianing Sun, Wei Guo, Dengcheng Zhang, Yingxue Zhang, Florence Regol, Yaochen Hu, Huifeng Guo, Ruiming Tang, Han Yuan, Xiuqiang He, et al. 2020. A Framework for Recommending Accurate and Diverse Items using Bayesian Graph Convolutional Neural Networks. In Proceedings of the 26...

    2020

  16. [24]

    Amos Tversky. 1969. Intransitivity of Preferences. Psychological review 76, 1 (1969), 31

    1969

  17. [25]

    Chenyang Wang, Yuanqing Yu, Weizhi Ma, Min Zhang, Chong Chen, Yiqun Liu, and Shaoping Ma. 2022. Towards Representation Alignment and Uniformity in Collaborative Filtering. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 1816–1825

    2022

  18. [26]

    Jun Wang, Lantao Yu, Weinan Zhang, Yu Gong, Yinghui Xu, Benyou Wang, Peng Zhang, and Dell Zhang. 2017. IRGAN: A Minimax Game for Unifying Generative and Discriminative Information Retrieval Models. In Proceedings of the 40th International ACM SIGIR Conference on Research and D...

    2017

  19. [27]

    Tongzhou Wang and Phillip Isola. 2020. Understanding Contrastive Represen- tation Learning through Alignment and Uniformity on the Hypersphere. In International Conference on Machine Learning . PMLR, 9929–9939

    2020

  20. [28]

    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 . 373–381

    2021

  21. [29]

    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

  22. [30]

    Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao, and Xinyi Le. 2022. Semi-Supervised Semantic Segmentation using Unreliable Pseudo-Labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 4248–4257

    2022

  23. [31]

    Jiancan Wu, Xiang Wang, Fuli Feng, Xiangnan He, Liang Chen, Jianxun Lian, and Xing Xie. 2021. Self-Supervised Graph Learning for Recommendation. In Proceed- ings of the 44th international ACM SIGIR conference on research and development in information retrieval. 726–735

    2021

  24. [32]

    Enneng Yang, Junwei Pan, Ximei Wang, Haibin Yu, Li Shen, Xihua Chen, Lei Xiao, Jie Jiang, and Guibing Guo. 2023. Adatask: A Task-Aware Adaptive Learning Rate Approach to Multi-Task Learning. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 10745–10753

    2023

  25. [33]

    Junliang Yu, Xin Xia, Tong Chen, Lizhen Cui, Nguyen Quoc Viet Hung, and Hongzhi Yin. 2023. XSimGCL: Towards Extremely Simple Graph Contrastive Learning for Recommendation. IEEE Transactions on Knowledge and Data Engi- neering (2023)

    2023

  26. [34]

    Junliang Yu, Hongzhi Yin, Min Gao, Xin Xia, Xiangliang Zhang, and Nguyen Quoc Viet Hung. 2021. Socially-Aware Self-Supervised Tri-training for Recommenda- tion. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2084–2092

    2021

  27. [35]

    Junliang Yu, Hongzhi Yin, Xin Xia, Tong Chen, Lizhen Cui, and Quoc Viet Hung Nguyen. 2022. Are Graph Augmentations Necessary? Simple Graph Contrastive Learning for Recommendation. InProceedings of the 45th International ACM SIGIR Conference on Research and Development in Infor...

    2022

  28. [36]

    Weinan Zhang, Tianqi Chen, Jun Wang, and Yong Yu. 2013. Optimizing Top-N Collaborative Filtering via Dynamic Negative Item Sampling. In Proceedings of the 36th International ACM SIGIR Conference on Research and Development in Information Retrieval. 785–788

    2013

  29. [37]

    Wayne Xin Zhao, Shanlei Mu, Yupeng Hou, Zihan Lin, Yushuo Chen, Xingyu Pan, Kaiyuan Li, Yujie Lu, Hui Wang, Changxin Tian, et al . 2021. Recbole: Towards a unified, comprehensive and efficient framework for recommendation algorithms. In Proceedings of the 30th ACM Internationa...

    2021

  30. [38]

    Yuhan Zhao, Rui Chen, Riwei Lai, Qilong Han, Hongtao Song, and Li Chen. 2023. Augmented Negative Sampling for Collaborative Filtering. In Proceedings of the 17th ACM Conference on Recommender Systems . 256–266

    2023

  31. [39]

    Qiannan Zhu, Haobo Zhang, Qing He, and Zhicheng Dou. 2022. A Gain-Tuning Dynamic Negative Sampler for Recommendation. In Proceedings of the ACM Web Conference. 277–285

    2022

  32. [40]

    Jianhuan Zhuo, Qiannan Zhu, Yinliang Yue, and Yuhong Zhao. 2022. Learning Explicit User Interest Boundary for Recommendation. In Proceedings of the ACM Web Conference. 193–202

    2022

  33. [2009]

    InProceedings of the 25th Conference on Uncertainty in Artificial Intelligence

    BPR: Bayesian Personalized Ranking from Implicit Feedback. InProceedings of the 25th Conference on Uncertainty in Artificial Intelligence . AUAI Press, 452– 461

  34. [2017]

    In Proceedings of the ACM Web Conference

    Neural Collaborative Filtering. In Proceedings of the ACM Web Conference . 173–182

  35. [2020]

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

    LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval . 639–648

Pith tools

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