Pith. sign in

REVIEW 4 major objections 4 minor 47 references

Contrastive Matrix Completion with Denoising and Augmented Graph Views for Robust Recommendation

T0 review · 4 major / 4 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 A contrastive graph model improves recommendation rankings by up to 36% while nudging rating accuracy.

desk verdict Plausible architecture and shipped code, but the sigmoid-vs-raw-ratings inconsistency makes the reported RMSE gains uninterpretable until the missing scaling is disclosed. read the letter →

arxiv 2506.10658 v1 pith:YREO6BUB submitted 2025-06-12 cs.IR cs.AIcs.NE

Narges Nemati , Mostafa Haghir Chehreghani This is my paper
classification cs.IRcs.AIcs.NE
keywords matrixcompletionrecommendersystemsgraphneuralnetworkscontrastivelearningdenoisingattentionmechanismvariationalautoencoder
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes a matrix-completion method, MCCL, that builds two views of each user–item interaction subgraph: one denoised by attention-weighted relational graph convolutions, and one generated by a variational graph autoencoder. The two views are trained together with a contrastive loss so they reinforce shared patterns, and their embeddings are combined for the final rating prediction. The authors claim this approach improves both rating accuracy (up to 0.8% RMSE reduction) and ranking quality (up to 36% gains in NDCG/MRR ranking metrics) on four Amazon datasets. The central idea is that explicitly attenuating noisy edges, rather than only random dropout, makes GNN-based recommendation more robust.

What carries the argument

The key machinery is the attention-weighted message-passing step in the denoising view: for each edge, the embeddings of its two endpoint nodes are concatenated and passed through an MLP, giving a raw attention score that is min–max normalized to [0,1] and then multiplied into the RGCN's per-relation message weights during propagation. This mechanism is what the paper claims removes noise. The second view uses a variational graph autoencoder (VGAE), which encodes each node as a Gaussian latent variable and adds a KL-divergence regularizer. A contrastive loss then pulls the two views' user and item embeddings into agreement, and the final prediction comes from a concatenation of both views' embeddings fed through an MLP.

What would settle it

Train the same model with random edge dropout (instead of learned attention weights) at matched capacity on the same four datasets; if random dropout achieves statistically equal or better RMSE and ranking metrics, the attention mechanism is not the source of the gain. Alternatively, inject known noisy edges into the test subgraphs and check whether low attention weights concentrate on those injected edges.

Watch

Extended reading notes

Core claim

The paper's central claim is that combining a denoising view and a variational autoencoder view through contrastive learning improves matrix completion for recommender systems. For the denoising view, an MLP computes an attention weight for each edge, the weights are normalized to [0,1], and these weights multiply the RGCN message-passing contributions, suppressing less important edges. For the second view, a variational graph autoencoder regularizes the latent distribution toward a standard prior. A mutual-learning contrastive loss aligns user and item embeddings across the two views, while separate losses preserve each view's individual strengths. The result, the paper reports on four Amazon datasets, is better RMSE than IGMC and KGMC and substantially better ranking metrics than all baselines, with the ranking gains far exceeding the numeric gains.

Load-bearing premise

The paper assumes that the learned attention weight for each edge, computed from endpoint embeddings and normalized to [0,1], actually identifies and attenuates noisy edges without discarding signal needed for rating prediction.

Editorial extensions

If this is right

  • If attention denoising works as claimed, GNN recommender models can be made robust to noisy edges without discarding rating information through random edge dropout.
  • The contrastive alignment between a denoised view and a variational view could serve as a general regularizer for inductive matrix completion, reducing overfitting on sparse interaction data.
  • Because the ranking improvements (up to 36%) are far larger than the RMSE improvements (up to 0.8%), the method is especially promising for top-N recommendation tasks where ranking quality matters more than exact rating values.
  • The loss decomposition into per-view losses plus a contrastive loss suggests a modular design that could be extended to other graph encoders beyond RGCN and VGAE.
  • The reported complexity stays linear in the number of edges, matching IGMC, so adding the contrastive and denoising machinery does not change the asymptotic cost.

Reading between the lines

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

  • A direct test of the denoising mechanism would be to inject synthetic noise into the interaction graph and measure whether the learned attention weights assign low values to the injected noisy edges; the paper does not report such a test.
  • The method's ranking gains might stem less from true denoising and more from the regularizing effect of the contrastive alignment with a variational view; comparing against random edge dropout with the same capacity would clarify this.
  • The ablation shows the denoising view alone already matches or beats the full model on several datasets, and on Amazon music it beats the full model, suggesting the VGAE view can sometimes hurt; a future extension could adaptively weight the views per dataset.
  • Because the paper tests only Amazon datasets, applying MCCL to implicit-feedback datasets or to non-rating interaction graphs would test whether the attention denoising transfers beyond explicit ratings.
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 / 4 minor

Summary. The paper proposes MCCL, a matrix-completion method for explicit-rating recommendation that extracts one-hop subgraphs around user-item interactions (following IGMC) and builds two views of each subgraph: an attention-denoised relational GCN view and a variational graph autoencoder view. The two views are trained with a combined loss that includes per-view MSE terms, a KL regularization term, a subgraph-reconstruction term, and a contrastive alignment loss. The authors evaluate on four Amazon datasets and report RMSE and ranking metrics against IGMC, SGL, SimGCL, and KGMC, claiming up to 0.8% RMSE improvement and up to 36% improvement in ranking metrics, and they include an ablation study and a hyperparameter study.

Significance. If the empirical claims were fully supported, the contribution would be a useful engineering combination of known components: subgraph extraction, relational message passing, attention-based edge reweighting, variational regularization, and contrastive alignment. The manuscript has some positive features: the code is publicly released, hyperparameters are tuned on a validation split, and the complexity analysis is included. However, the central claims are not currently supported as stated. The final prediction head appears to be inconsistent with the training targets, the ranking claim is contradicted by the office-dataset results in Table 3, no uncertainty quantification is provided for very small RMSE differences, and the ablation does not establish the claimed denoising and mutual-learning mechanisms. These issues are load-bearing for the paper's main message and require substantial revision.

major comments (4)
  1. [4.2.5, Algorithm 3, Eq. (12)] The final prediction head applies a sigmoid to the output of the MLP (Algorithm 3, line 18), while all prediction losses, including L_pred_final in Eq. (12), are computed as MSE against raw 1-5 ratings. Since a sigmoid output is strictly in (0,1), the MSE loss against 1-5 targets would be minimized by predicting near a clipped conditional mean, and it cannot produce the reported RMSE values around 0.85-0.96. The manuscript never states that ratings are normalized or that the sigmoid output is rescaled. This is an internal inconsistency: every improvement percentage in Tables 2-5 depends on the missing scaling or the missing change of output activation. Please specify the exact rating preprocessing and output transformation, or change the prediction head and rerun the experiments.
  2. [5.5, Table 3] The abstract and Section 5.5 claim that MCCL produces 'superior rankings,' but Table 3 shows that on the Amazon office dataset MCCL is worse than SimGCL on all four ranking metrics (e.g., NDCG@10-ranking 0.234 vs. 0.249 and MRR@10-ranking 0.210 vs. 0.222), and the improvement row reports -5%, -4%, -6%, -6%. The authors do explicitly mention SimGCL's better office ranking performance in the text, but the abstract and conclusion present the ranking improvement as a general result. The claim should be qualified as dataset-dependent, and the improvement row should state explicitly which baseline the percentages are computed against.
  3. [5.4, 5.5] All results in Tables 2-5 are reported as single numbers with no repeated runs, no standard deviations, and no significance tests. Because the RMSE improvements are very small (0.4%-0.7%), it is not clear whether these differences are stable or within run-to-run noise. At minimum, report means and standard deviations over multiple random seeds and, preferably, a paired significance test for the RMSE and ranking metrics.
  4. [4.2.3, Table 6] The paper claims that the attention mechanism removes noisy edges, but no experiment shows that low-attention edges are actually noisy, and there is no comparison against random edge dropout with the same capacity. Moreover, the ablation in Table 6 shows that the denoising component alone matches or outperforms the full MCCL model on several datasets (e.g., on Amazon music, denoising has NDCG@10-ranking 0.301 vs. MCCL's 0.284; on tools, both have RMSE 0.963). This is acknowledged in the text, but it undermines the claim that the contrastive combination of the two views is the source of the improvement. Additional experiments or a revised interpretation are needed.
minor comments (4)
  1. [6.1.1, 6.1.2, 6.1.3] The figure numbering and dataset references are inconsistent: Section 6.1.1 says the optimal alpha for Amazon movie is 0.001 and for Amazon music is 0.0045, but the accompanying text and captions mix up movie, office, and music; Section 6.1.2 refers to 'Figure 5' for the beta study, which appears to be Figure 4, and Section 6.1.3 also refers to Figure 5. Please renumber the figures and align the captions with the datasets actually used.
  2. [Eq. (7)] There is a typographical error in Eq. (7): '-D_KL(q(z|Ap(z))' is missing a closing parenthesis and the notation for the KL divergence should be written as D_KL(q(z|A) || p(z)) to be consistent with Eq. (8).
  3. [Tables 2-5] The 'Improvement' rows should state the reference baseline explicitly (e.g., 'relative to the best baseline' or 'relative to IGMC'), since on the office dataset the values are negative while the RMSE improvement over IGMC is positive; the current format is ambiguous.
  4. [5.2] The baseline descriptions for SGL and SimGCL say they use BPR loss and are designed for implicit feedback, but they are then evaluated on an explicit-rating RMSE task. This is not necessarily a flaw, but the paper should acknowledge that RMSE comparisons with implicit-feedback baselines may be less informative than ranking comparisons.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MCCL is an empirical pipeline of standard views and losses, with no equation reducing to fitted values and no load-bearing self-citation chain.

full rationale

The paper's claimed derivation is not circular. The method takes IGMC-style extracted subgraphs, builds two views (an attention-weighted RGCN denoising view and a VGAE view), trains with MSE losses (Eqs. 4, 5, 6, 12), KL regularization (Eqs. 8-9), and contrastive InfoNCE losses (Eq. 11), then combines view embeddings through an MLP+sigmoid head (Algorithm 3, lines 16-18). None of these equations define the final prediction in terms of the reported RMSE or ranking improvements, and no fitted parameter is renamed as a prediction: the prediction heads are trained by the same MSE objective that is evaluated, which is standard supervised learning rather than circularity. Hyperparameters (alpha, beta, lambda) are tuned on a validation split (Section 5.4) and then applied to test data, so the reported gains are not forced by construction. The self-citations ([1], [7], [12]-[14], [47]) appear in preliminaries and related work and supply no load-bearing uniqueness claim or ansatz; [1] is described only as related dynamic recommendation work. The acknowledged cases where MCCL does not win (Amazon office ranking versus SimGCL, and Amazon music where the denoising component alone is better) are limitation/consistency issues, and the sigmoid-versus-raw-rating-scale concern is a correctness or documentation risk, but neither is a reduction of the central claim to its own inputs. No circular step is present.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on two domain assumptions the paper does not test directly: that 1-hop subgraphs are sufficient for rating prediction, and that learned attention weights identify genuinely noisy edges. The rest is standard VAE and GNN machinery. The tunable loss weights and unreported temperature are the main fitted quantities.

free parameters (6)
  • alpha (L_rec weight) = 0.001 on movie, 0.0045 on music (text inconsistent)
    Controls reconstruction loss in the VGAE view. Tuned on validation; the paper lists tested values [0.01, 0.0045, 0.001] and gives different optimal values for different datasets, with office/music confusion in the text.
  • beta (L_KL weight) = 0.001 (movie)
    Weights the KL divergence term in the VGAE loss. Tuned on validation values [0.01, 0.0045, 0.001, 0.0001]; only the movie optimal is reported.
  • lambda (contrastive loss weight) = 0.001 (movie)
    Weights the InfoNCE contrastive loss in the final loss. Tuned on validation values [0.003, 0.002, 0.001, 0.0001]; only the movie optimal is reported.
  • tau (contrastive temperature) = not reported
    Temperature in the contrastive loss (Eq. 11). The paper never states its value or tuning range, which is a reproducibility gap.
  • initial one-hot node embeddings = [1,0,0,0], [0,1,0,0], [0,0,1,0], [0,0,0,1]
    Hand-designed initial features for target user, target item, neighbor user, and neighbor item, following IGMC (Section 4.2.1). This encoding is chosen by hand and is not learned.
  • learning rate, weight decay, batch size = 0.001, 0.09, 128
    Fixed across experiments; learning rate is shared with baselines, weight decay is specific to MCCL (Section 5.4). These are tuned choices not derived from theory.
assumptions (4)
  • domain assumption Local neighborhood subgraphs around a target user-item pair contain enough signal to predict the missing rating.
    The entire method inherits IGMC's inductive bias that a 1-hop subgraph of the bipartite graph is sufficient for rating prediction. Section 4.2.1.
  • domain assumption Noisy edges can be identified from concatenated endpoint embeddings by an MLP attention module and attenuated without removing useful signal.
    Section 4.2.3, Eq. (3). The paper does not test this assumption directly; it only measures end-to-end metrics.
  • standard math The reparameterization trick and KL divergence for the variational graph autoencoder are valid and applied correctly.
    Section 4.2.4, Eq. (8), standard VAE machinery from [22, 23].
  • domain assumption Five-core filtered Amazon datasets are representative of recommendation settings and the 60/20/20 split is unbiased.
    Section 5.3, Section 5.4. No analysis of how filtering affects conclusions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Contrastive Matrix Completion with Denoising and Augmented Graph Views for Robust Recommendation." pith.science (2026). https://pith.science/paper/YREO6BUB

@misc{pith2026250610658,
  author       = {Pith},
  title        = {Pith review of: Contrastive Matrix Completion with Denoising and Augmented Graph Views for Robust Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YREO6BUB}},
  note         = {Machine review of arXiv:2506.10658}
}
read the original abstract

Matrix completion is a widely adopted framework in recommender systems, as predicting the missing entries in the user-item rating matrix enables a comprehensive understanding of user preferences. However, current graph neural network (GNN)-based approaches are highly sensitive to noisy or irrelevant edges--due to their inherent message-passing mechanisms--and are prone to overfitting, which limits their generalizability. To overcome these challenges, we propose a novel method called Matrix Completion using Contrastive Learning (MCCL). Our approach begins by extracting local neighborhood subgraphs for each interaction and subsequently generates two distinct graph representations. The first representation emphasizes denoising by integrating GNN layers with an attention mechanism, while the second is obtained via a graph variational autoencoder that aligns the feature distribution with a standard prior. A mutual learning loss function is employed during training to gradually harmonize these representations, enabling the model to capture common patterns and significantly enhance its generalizability. Extensive experiments on several real-world datasets demonstrate that our approach not only improves the numerical accuracy of the predicted scores--achieving up to a 0.8% improvement in RMSE--but also produces superior rankings with improvements of up to 36% in ranking metrics.

Figures

Figures reproduced from arXiv: 2506.10658 by the authors.

Figure 1
Figure 1. Overall framework of the proposed MCCL model. [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Impact of the α Parameter on the Amazon movie Dataset. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_2.png] view at source ↗
Figure 3
Figure 3. Impact of the α Parameter on the Amazon music Dataset. 6.1.2 Effect of the β coefficient To evaluate the impact of the β coefficient on the Amazon movie dataset, we test its different values: [0.01, 0.0045, 0.001, 0.0001]. The results are presented in [PITH_FULL_IMAGE:figures/full_fig_p024_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Impact of the β Parameter on the Amazon movie Dataset. 6.1.3 Effect of the λ coefficient To analyze the impact of the λ coefficient on the Amazon movie dataset, we test its different values: [0.003, 0.002, 0.001, 0.0001]. The results are presented in [PITH_FULL_IMAGE:…
Figure 5
Figure 5. Figure 5: Impact of the λ Parameter on the Amazon movie Dataset. 7 Conclusion In this paper, we addressed the problem of matrix completion in recommender systems. Our proposed method first extracted local subgraphs for each interaction. We then employed a deep learning framework…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 42 canonical work pages

  1. [1]

    Adaptive long-term embedding with denoising and augmentation for recommendation, 2025

    Zahra Akhlaghi and Mostafa Haghir Chehreghani. Adaptive long-term embedding with denoising and augmentation for recommendation, 2025. 25

    work page 2025

  2. [2]

    A systematic study on the recommender systems in the e- commerce.IEEE Access, 8:115694–115716, 2020

    Pegah Malekpour Alamdari, Nima Jafari Navimipour, Mehdi Hosseinzadeh, Ali Asghar Safaei, and Aso Mohammad Darwesh. A systematic study on the recommender systems in the e- commerce.IEEE Access, 8:115694–115716, 2020

    work page 2020

  3. [3]

    Facial emotion recognition and music recommendation system using cnn-based deep learning techniques.Evol

    Brijesh Bakariya, Arshdeep Singh, Harmanpreet Singh, Pankaj Raju, Rohit Rajpoot, and Kr- ishna Kumar Mohbey. Facial emotion recognition and music recommendation system using cnn-based deep learning techniques.Evol. Syst., 15(2):641–658, 2024

    work page 2024

  4. [4]

    Cand` es, and Zuowei Shen

    Jian-Feng Cai, Emmanuel J. Cand` es, and Zuowei Shen. A singular value thresholding algorithm for matrix completion.SIAM J. Optim., 20(4):1956–1982, 2010

    work page 1956

  5. [5]

    Lightgcl: Simple yet effective graph contrastive learning for recommendation

    Xuheng Cai, Chao Huang, Lianghao Xia, and Xubin Ren. Lightgcl: Simple yet effective graph contrastive learning for recommendation. InThe Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023

    work page 2023

  6. [6]

    Cand` es and Benjamin Recht

    Emmanuel J. Cand` es and Benjamin Recht. Exact matrix completion via convex optimization. Commun. ACM, 55(6):111–119, 2012

    work page 2012

  7. [7]

    Half a decade of graph convolutional networks.Nat

    Mostafa Haghir Chehreghani. Half a decade of graph convolutional networks.Nat. Mach. Intell., 4(3):192–193, 2022

    work page 2022

  8. [8]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple frame- work for contrastive learning of visual representations. InProceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 1597–1607. PMLR, 2020

    work page 2020

Show all 47 references
  1. [9]

    Probabilistic ma- trix factorization recommendation approach for integrating multiple information sources.IEEE Trans

    Jiangzhou Deng, Xun Ran, Yong Wang, Leo Yu Zhang, and Junpeng Guo. Probabilistic ma- trix factorization recommendation approach for integrating multiple information sources.IEEE Trans. Syst. Man Cybern. Syst., 53(10):6220–6231, 2023

    2023

  2. [10]

    Geometric matrix completion via sylvester multi-graph neural network

    Boxin Du, Changhe Yuan, Fei Wang, and Hanghang Tong. Geometric matrix completion via sylvester multi-graph neural network. In Ingo Frommholz, Frank Hopfgartner, Mark Lee, Michael Oakes, Mounia Lalmas, Min Zhang, and Rodrygo L. T. Santos, editors,Proceedings of the 32nd ACM Int...

    2023

  3. [11]

    A survey of explainable e-commerce recommender systems

    Huaqi Gao and Shunke Zhou. A survey of explainable e-commerce recommender systems. In IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology, WI-IAT 2022, Niagara Falls, ON, Canada, November 17-20, 2022, pages 723–730. IEEE, 2022

    2022

  4. [12]

    Content augmented graph neural networks.ACM Trans

    Fatemeh Gholamzadeh Nasrabadi, Amirhossein Kashani, Pegah Zahedi, and Mostafa Haghir Chehreghani. Content augmented graph neural networks.ACM Trans. Web, October

  5. [13]

    Heterophily-aware fair recommendation using graph convolutional networks.CoRR, abs/2402.03365, 2024

    Nemat Gholinejad and Mostafa Haghir Chehreghani. Heterophily-aware fair recommendation using graph convolutional networks.CoRR, abs/2402.03365, 2024

    2024 arXiv

  6. [14]

    Disentangling popularity and quality: An edge classification approach for fair recommendation.CoRR, abs/2502.15699, 2025

    Nemat Gholinejad and Mostafa Haghir Chehreghani. Disentangling popularity and quality: An edge classification approach for fair recommendation.CoRR, abs/2502.15699, 2025

    2025 arXiv

  7. [15]

    Keyword-enhanced recommender system based on inductive graph matrix completion.Eng

    Donghee Han, Daehee Kim, Keejun Han, and Mun Yong Yi. Keyword-enhanced recommender system based on inductive graph matrix completion.Eng. Appl. Artif. Intell., 128:107499, 2024

    2024

  8. [16]

    Neural collab- orative filtering

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. Neural collab- orative filtering. In Rick Barrett, Rick Cummings, Eugene Agichtein, and Evgeniy Gabrilovich, editors,Proceedings of the 26th International Conference on World Wide Web, WWW 2017, Per...

    2017

  9. [17]

    Reducing the dimensionality of data with neural networks.science, 313(5786):504–507, 2006

    Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks.science, 313(5786):504–507, 2006

    2006

  10. [18]

    imovierec: a hybrid movie recommendation method based on a user-image-item model.Int

    Syjung Hwang, Hyeongjin Ahn, and Eunil Park. imovierec: a hybrid movie recommendation method based on a user-image-item model.Int. J. Mach. Learn. Cybern., 14(9):3205–3216, 2023

    2023

  11. [19]

    Regularized singular value decomposition in news recommendation system

    Youchun Ji, Wenxing Hong, Yali Shangguan, Huan Wang, and Jing Ma. Regularized singular value decomposition in news recommendation system. In11th International Conference on Com- puter Science & Education, ICCSE 2016, Nagoya, Japan, August 23-25, 2016, pages 621–626. IEEE, 2016

    2016

  12. [20]

    Adaptive graph contrastive learning for recommendation

    Yangqin Jiang, Chao Huang, and Lianghao Huang. Adaptive graph contrastive learning for recommendation. In Ambuj K. Singh, Yizhou Sun, Leman Akoglu, Dimitrios Gunopulos, Xifeng Yan, Ravi Kumar, Fatma Ozcan, and Jieping Ye, editors,Proceedings of the 29th ACM SIGKDD Conference o...

    2023

  13. [21]

    Recent developments in affective recommender systems

    Rahul Katarya and Om Prakash Verma. Recent developments in affective recommender systems. Physica A: Statistical Mechanics and its Applications, 461:182–190, 2016

    2016

  14. [22]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In Yoshua Bengio and Yann LeCun, editors,2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings, 2014

    2014

  15. [23]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Variational graph auto-encoders.CoRR, abs/1611.07308, 2016

    2016 arXiv

  16. [24]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional net- works. In5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. 27

    2017

  17. [25]

    Factorization meets the neighborhood: a multifaceted collaborative filtering model

    Yehuda Koren. Factorization meets the neighborhood: a multifaceted collaborative filtering model. In Ying Li, Bing Liu, and Sunita Sarawagi, editors,Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Las Vegas, Nevada, USA, Augu...

    2008

  18. [26]

    Bell, and Chris Volinsky

    Yehuda Koren, Robert M. Bell, and Chris Volinsky. Matrix factorization techniques for recom- mender systems.Computer, 42(8):30–37, 2009

    2009

  19. [27]

    Lee and H

    Daniel D. Lee and H. Sebastian Seung. Algorithms for non-negative matrix factorization. In Todd K. Leen, Thomas G. Dietterich, and Volker Tresp, editors,Advances in Neural Informa- tion Processing Systems 13, Papers from Neural Information Processing Systems (NIPS) 2000, Denve...

    2000

  20. [28]

    Spectral regularization algorithms for learning large incomplete matrices.J

    Rahul Mazumder, Trevor Hastie, and Robert Tibshirani. Spectral regularization algorithms for learning large incomplete matrices.J. Mach. Learn. Res., 11:2287–2322, 2010

    2010

  21. [29]

    Intelligible graph contrastive learning with attention-aware for recommendation.Neurocomputing, 614:128781, 2025

    Xian Mo, Zihang Zhao, Xiaoru He, Hang Qi, and Hao Liu. Intelligible graph contrastive learning with attention-aware for recommendation.Neurocomputing, 614:128781, 2025

    2025

  22. [30]

    Current challenges and visions in music recommender systems research.Int

    Markus Schedl, Hamed Zamani, Ching-Wei Chen, Yashar Deldjoo, and Mehdi Elahi. Current challenges and visions in music recommender systems research.Int. J. Multim. Inf. Retr., 7(2):95– 116, 2018

    2018

  23. [31]

    Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling

    Michael Sejr Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling. Modeling relational data with graph convolutional networks. In Aldo Gangemi, Roberto Navigli, Maria-Esther Vidal, Pascal Hitzler, Rapha¨ el Troncy, Laura Hollink, Anna To...

    2018

  24. [32]

    Autorec: Autoencoders meet collaborative filtering

    Suvash Sedhain, Aditya Krishna Menon, Scott Sanner, and Lexing Xie. Autorec: Autoencoders meet collaborative filtering. In Aldo Gangemi, Stefano Leonardi, and Alessandro Panconesi, ed- itors,Proceedings of the 24th International Conference on World Wide Web Companion, WWW 2015...

    2015

  25. [33]

    Inductive matrix completion using graph autoencoder

    Wei Shen, Chuheng Zhang, Yun Tian, Liang Zeng, Xiaonan He, Wanchun Dou, and Xiaolong Xu. Inductive matrix completion using graph autoencoder. In Gianluca Demartini, Guido Zuccon, J. Shane Culpepper, Zi Huang, and Hanghang Tong, editors,CIKM ’21: The 30th ACM Inter- national Co...

    2021

  26. [34]

    Context-aware recommender systems for social networks: Review, challenges and opportunities.IEEE Access, 9:57440–57463, 2021

    Areej Bin Suhaim and Jawad Berri. Context-aware recommender systems for social networks: Review, challenges and opportunities.IEEE Access, 9:57440–57463, 2021. 28

    2021

  27. [35]

    Kipf, and Max Welling

    Rianne van den Berg, Thomas N. Kipf, and Max Welling. Graph convolutional matrix comple- tion.CoRR, abs/1706.02263, 2017

    2017 arXiv

  28. [36]

    Theoretically accurate regularization technique for matrix factorization based rec- ommender systems

    Hao Wang. Theoretically accurate regularization technique for matrix factorization based rec- ommender systems. InCSAI 2021: 5th International Conference on Computer Science and Artificial Intelligence, Beijing, China, December 4 - 6, 2021, pages 256–259. ACM, 2021

    2021

  29. [37]

    Multi-level cross-modal con- trastive learning for review-aware recommendation.Expert Syst

    Yibiao Wei, Yang Xu, Lei Zhu, Jingwei Ma, and Chengmei Peng. Multi-level cross-modal con- trastive learning for review-aware recommendation.Expert Syst. Appl., 247:123341, 2024

    2024

  30. [38]

    Solving a low-rank factorization model for matrix com- pletion by a nonlinear successive over-relaxation algorithm.Math

    Zaiwen Wen, Wotao Yin, and Yin Zhang. Solving a low-rank factorization model for matrix com- pletion by a nonlinear successive over-relaxation algorithm.Math. Program. Comput., 4(4):333– 361, 2012

    2012

  31. [39]

    Self-supervised graph learning for recommendation

    Jiancan Wu, Xiang Wang, Fuli Feng, Xiangnan He, Liang Chen, Jianxun Lian, and Xing Xie. Self-supervised graph learning for recommendation. In Fernando Diaz, Chirag Shah, Torsten Suel, Pablo Castells, Rosie Jones, and Tetsuya Sakai, editors,SIGIR ’21: The 44th International ACM...

    2021

  32. [40]

    A neural matrix decomposition recommender system model based on the multimodal large language model

    Ao Xiang, Bingjie Huang, Xinyu Guo, Haowei Yang, and Tianyao Zheng. A neural matrix decomposition recommender system model based on the multimodal large language model. In Proceedings of the 2024 7th International Conference on Machine Learning and Machine Intel- ligence (MLMI...

    2024

  33. [41]

    Are graph augmentations necessary?: Simple graph contrastive learning for recommendation

    Junliang Yu, Hongzhi Yin, Xin Xia, Tong Chen, Lizhen Cui, and Quoc Viet Hung Nguyen. Are graph augmentations necessary?: Simple graph contrastive learning for recommendation. In Enrique Amig´ o, Pablo Castells, Julio Gonzalo, Ben Carterette, J. Shane Culpepper, and Gabriella K...

    2022

  34. [42]

    Inductive matrix completion based on graph neural networks

    Muhan Zhang and Yixin Chen. Inductive matrix completion based on graph neural networks. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net, 2020

    2020

  35. [43]

    Wilkinson, Robert Schreiber, and Rong Pan

    Yunhong Zhou, Dennis M. Wilkinson, Robert Schreiber, and Rong Pan. Large-scale parallel collaborative filtering for the netflix prize. In Rudolf Fleischer and Jinhui Xu, editors,Algorithmic Aspects in Information and Management, 4th International Conference, AAIM 2008, Shangha...

    2008

  36. [44]

    Bayesian generalized low rank regression models for neuroimaging phenotypes and genetic markers.Journal of the American Statistical Association, 109(507):977–990, 2014

    Hongtu Zhu, Zakaria Khondker, Zhaohua Lu, and Joseph G Ibrahim. Bayesian generalized low rank regression models for neuroimaging phenotypes and genetic markers.Journal of the American Statistical Association, 109(507):977–990, 2014

    2014

  37. [45]

    Matrix completion of adaptive jumping graph neural networks for recommendation systems.IEEE Access, 11:88433–88450, 2023

    Xiaodong Zhu, Junyu Fu, and Chen Chen. Matrix completion of adaptive jumping graph neural networks for recommendation systems.IEEE Access, 11:88433–88450, 2023

    2023

  38. [46]

    Sina Ziaee, Hossein Rahmani, and Mohammad Nazari

    S. Sina Ziaee, Hossein Rahmani, and Mohammad Nazari. Morgh: movie recommender system using gnns on heterogeneous graphs.Knowl. Inf. Syst., 66(12):7419–7435, 2024

    2024

  39. [47]

    Centrality- based and similarity-based neighborhood extension in graph neural networks.J

    Mohammadjavad Zohrabi, Saeed Saravani, and Mostafa Haghir Chehreghani. Centrality- based and similarity-based neighborhood extension in graph neural networks.J. Supercomput., 80(16):24638–24663, 2024. 30

    2024

Pith tools

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