Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

NLGCL: Naturally Existing Neighbor Layers Graph Contrastive Learning for Recommendation

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that graph contrastive learning for recommendation can be built from the naturally existing views between adjacent GNN layers, without augmentation, and reports consistent gains in accuracy and speed on four public…

desk verdict A genuinely useful augmentation-free contrastive view idea with believable empirical gains, but the appendix proof of Theorem 1 is invalid and the first-G choice is currently ungrounded; worth a serious referee only if the authors fix or drop that theorem. read the letter →

arxiv 2507.07522 v1 pith:ERXMV33L submitted 2025-07-10 cs.IR

Jinfeng Xu , Zheyu Chen , Shuo Yang , Jinze Li , Hewei Wang , Wei Wang , Xiping Hu , Edith Ngai This is my paper · ORCID
classification cs.IR
keywords graphcontrastivelearningrecommendersystemscollaborativefilteringneuralnetworksneighbor-layerviewsself-supervisedLightGCNInfoNCE
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

Graph contrastive learning for recommender systems normally pays for augmentation: it corrupts graphs or embeddings to make contrastive views, which adds noise and many extra operations. This paper argues that a GNN already contains natural contrastive views: the embeddings of a node at one layer and the embeddings of its neighbors in the next layer. The proposed objective, NLGCL, treats each node and its neighbors in the next layer as positive pairs and all other nodes as negatives, requiring no augmentation and no extra view storage. On Yelp, Pinterest, QB-Video, and Alibaba, the paper reports that this simple objective beats the compared state-of-the-art baselines on Recall and NDCG while reducing total training time. If correct, it means the main cost and a main source of noise in graph contrastive recommenders can simply be removed.

What carries the argument

The central object is the neighbor-layer contrastive pair: for each user $u$ at layer $l-1$, the positive set is $\{e^{(l)}_{\tilde i}: \tilde i \in \mathcal N_u\}$, the next-layer embeddings of items $u$ interacted with, and symmetrically for items. Message passing makes each next-layer neighbor embedding a weighted aggregation that includes the anchor node itself, so the pair is semantically linked rather than randomly augmented. The final objective adds this neighbor-layer InfoNCE-style loss to the BPR ranking loss and an L2 regularizer, with a hyperparameter $G$ controlling how many of the $L$ layer groups are used. This mechanism carries the argument by supplying the positive signal for free: no augmented graph, no noise injection, and no extra view storage.

What would settle it

Estimate the mutual information (or a practical proxy such as the cosine similarity between layer-$l$ and layer-$(l+1)$ embeddings) on Yelp and Pinterest and check whether it decreases monotonically as $l$ grows; if it does not, Theorem 1 is false. A second check: train NLGCL with $G$ chosen from layers 2 through 4 instead of the first $G$ layers; if later-layer views ever match or beat the first-$G$ choice, the paper's central design rule is falsified.

Watch

Extended reading notes

Core claim

The paper claims that adjacent layers of a LightGCN-style encoder form contrastive views naturally: layer $l-1$ embeddings of a user or item are positively paired with layer $l$ embeddings of its neighbors, because those neighbor embeddings are aggregated from the same node and semantically similar users or items. All other nodes serve as negatives, either restricted to same-type nodes in the heterogeneous scope or spanning all users and items in the entire scope. The contrastive loss is averaged over the first $G$ layer groups, a selection justified by a theorem asserting that earlier-layer views carry more mutual information. On four public datasets, the method is reported to outperform all compared baselines on Recall@K and NDCG@K, including augmentation-based contrastive learning methods, while converging in less wall-clock time.

Load-bearing premise

The argument that only the first $G$ layer groups should be used depends on the theorem that earlier-layer contrastive views are always more informative; if that theorem does not hold, the layer-selection rule is unproven, though the empirical method could still be useful.

Editorial extensions

If this is right

  • Graph contrastive learning for collaborative filtering can be defined without any augmentation operator; the encoder's own layer outputs already supply the positives and negatives.
  • Training cost drops substantially: on Yelp, the heterogeneous-scope NLGCL converged in 49 minutes with NDCG@10 of 0.0713, while the strongest augmentation-based competitor took over four hours.
  • Treating all nodes as negatives in the entire scope is consistently worse than restricting negatives to same-type nodes in the heterogeneous scope, so cross-type negatives add semantic noise rather than signal.
  • The number $G$ of layer groups acts as a direct accuracy-efficiency dial: $G=2$ was optimal on all four datasets, and even $G=1$ already exceeded every baseline in the comparison.

Reading between the lines

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

  • Beyond the paper, the same neighbor-layer objective should transfer to any GNN-based encoder, not just the LightGCN backbone, because the positive pairs are read off the forward pass; this is a testable extension the authors do not run.
  • The reported uniformity of learned embeddings suggests a direct test of popularity-bias reduction: compare long-tail item Recall for NLGCL against SimGCL on the same four datasets.
  • If Theorem 1 does not survive scrutiny, an alternative design is to learn per-layer-group weights rather than committing to the first $G$ groups, which could adapt to datasets of different densities.
  • Validating the optimal $G=2$ on a held-out split or additional datasets would clarify whether the layer-group count is a stable property of the approach or a per-dataset tuning artifact.
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

3 major / 5 minor

Summary. The paper proposes NLGCL, a graph contrastive learning method for collaborative filtering that avoids explicit data augmentation by using the embeddings produced at adjacent GNN layers as naturally existing contrastive views. Positive pairs are formed between a node at layer l-1 and its neighbors at layer l, while negatives are other nodes in the layer-l view; a heterogeneous and an entire scope are defined. The method is evaluated on four datasets and is reported to outperform state-of-the-art CF and CL baselines in both ranking metrics and training efficiency. The authors also claim a theoretical proof, in Theorem 1 of Appendix A.1, that the first G layer groups are the optimal choice for contrastive learning.

Significance. If the empirical results hold, the paper makes a useful practical contribution: an augmentation-free contrastive objective that reuses GNN layer outputs, reducing the storage and computation overhead of view construction while improving accuracy. The experiments are broad (four datasets, eleven baselines), the improvements are consistent, and the authors provide code. The main weakness is that the theoretical justification for the architecture choice, Theorem 1, is invalid as stated. The empirical claims may still stand, but the paper currently overstates its theoretical support. The central method is simple and plausible, so the contribution is worth considering after a major revision that repairs or removes the faulty proof and replaces it with direct evidence for the first-G design choice.

major comments (3)
  1. [Section 4.2 / Appendix A.1] The proof of Theorem 1 is invalid, and the theorem is load-bearing because Section 4.2 states that the first G groups are optimal 'by theoretical proof.' In Eq. (13), the Gaussian mutual information formula is written as 1/2 log(|Σ(l-1)|/|Σ(l)|), but for jointly Gaussian variables the correct expression is 1/2 log(|Σ_X||Σ_Y|/|Σ_XY|). For the deterministic map E(l) = ÃE(l-1), Eq. (14) substitutes Σ(l) = òΣ(l-1), whereas the covariance update is Σ(l) = ÃΣ(l-1)Ãᵀ; the simplification assumes commutativity and ignores the cross-covariance term. Eq. (15) then yields an expression independent of l, directly contradicting the claimed O(λ_max^{2l}) decay. The theorem's conclusion therefore has no valid derivation, and the architecture choice G ≤ L and the associated efficiency argument are not theoretically grounded.
  2. [Appendix A.1, Lemma 1] The entropy-reduction lemma is not proven as stated. The proof invokes the entropy-power inequality, which applies to sums of independent random variables under additive noise, but the propagation E(l) = ÃE(l-1) is a deterministic linear transformation. The inequality does not apply in this setting without additional assumptions, and the covariance update used in the proof is again incorrect because Cov(ÃE) = ÃCov(E)Ãᵀ, not òCov(E). The lemma needs a correct proof or should be removed.
  3. [Section 4.2 / Theorem 1] Even if a correct statement that I(E(l-1);E(l)) decreases with l were available, it would not establish that lower-layer contrastive pairs are 'more effective' for contrastive learning. The proof asserts without derivation that the signal-to-noise ratio is proportional to this mutual information and that this translates directly to recommendation performance. This missing link is central to the claim that the first G groups are optimal. The paper should either provide a rigorous connection between layerwise mutual information and the InfoNCE objective or replace the theoretical claim with an empirical study. Figure 6 only varies G for the first G groups; it does not compare first-G groups against last-G or randomly selected groups, so the claimed optimality of the first groups is not empirically established either.
minor comments (5)
  1. [Section 5.2, Observation1] The quoted NDCG@10 improvements do not match Table 3: for Pinterest the 5.69% value corresponds to NDCG@20, for QB-Video the 2.52% value is NDCG@20, and for Alibaba the 12.11% value is NDCG@20; the text should report the corresponding NDCG@10 values or clarify which metric is used.
  2. [Table 4, Alibaba block] The N@10 values listed for NCL (0.0477) and LightGCL (0.0459) are identical to their Recall@10 values in Table 3, while the NDCG@10 values in Table 3 are 0.0259 and 0.0239, respectively; the table appears to have mixed in Recall values, so the efficiency comparison needs to be corrected.
  3. [Figure 3] The second subplot is labeled 'Time on QB-Video/minutes' but the text and the surrounding figure panels indicate that the second dataset is Pinterest; the axis labels and panel ordering should be checked and corrected.
  4. [Equations (6)-(9)] The use of a product symbol inside the logarithm for the positive-pair terms is unusual for an InfoNCE-style loss; if a sum over positive samples is intended, the notation should be changed to a sum, and the normalization by 1/|N_u| should be explained.
  5. [Section 2.1] The notation R+ ∈ {R_u,i | u ∈ U, i ∈ I, R_u,i = 1} is ambiguous because R+ is described as a matrix but the set contains scalar entries; it would be clearer to define the sets of observed and unobserved interaction entries directly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the CL objective is trained jointly with BPR, no fitted value is renamed as a prediction, and self-citations are related-work only.

full rationale

NLGCL's central empirical claim (outperforming baselines on four datasets) is established by held-out evaluation against external benchmarks, not by construction from its own loss. The contrastive loss is a regularizer trained jointly with the BPR loss, so there is no fitted parameter that is later 'predicted.' The positive-pair definition (a node with its next-layer neighbors) is a structural choice derived from GCN message passing rather than a quantity fitted to the target metric; the statement that such pairs are semantically relevant follows from the aggregation equation but does not by itself force the reported accuracy improvements. Related-work citations to the authors' own prior papers are contextual and not load-bearing for the method's correctness. The Appendix A.1 proof of Theorem 1 contains mathematical errors (the Gaussian MI formula and the l-independent expression in Eq. 15), but an invalid proof is a correctness or rigor issue, not circularity: Theorem 1 is not assumed as an input, and the paper does not cite an external uniqueness theorem from the authors to forbid alternatives. The efficiency analysis is a complexity comparison, and the empirical efficiency results are measured. Hence no step in the derivation chain reduces, by definition or by self-citation, to its own inputs.

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

The model itself adds no new physical entities. The main free parameters are standard hyperparameters tuned on validation. The critical axioms are the unsupported entropy-reduction lemma used in the theorem proof and the domain assumption that interaction-neighbor semantics are shared. No invented entities are introduced.

free parameters (6)
  • lambda_1 (CL loss weight) = 1e-5 (optimal on all datasets)
    Tuned over {1e-6, 1e-5, 1e-4}; chosen value reported in Section 5.6.
  • tau (temperature) = 0.2
    Tuned over {0.1, 0.2, 0.3, 0.4}; chosen value reported in Section 5.6.
  • L (number of GNN layers) = 2 (Yelp, Alibaba), 3 (Pinterest, QB-Video)
    Tuned over {1,2,3,4}; optimal per dataset in Figure 6.
  • G (number of contrastive view groups) = 2 (all datasets)
    Tuned over {1,...,L}; optimal value claimed in Section 5.6 and Figure 6.
  • lambda_2 (L2 regularization weight) = 1e-4
    Fixed, not tuned; mentioned in Section 5.1.3.
  • embedding size d = 64
    Fixed model dimension; a design choice, not fitted to data.
assumptions (4)
  • ad hoc to paper Entropy reduction via low-pass filtering: propagation E^(l) = A_tilde E^(l-1) reduces entropy, i.e., H(E^(l)) <= H(E^(l-1)).
    Invoked in Appendix A.1 as Lemma 1 to prove Theorem 1; the proof is flawed because the covariance substitution and mutual information decay derivation are incorrect. This assertion is used to justify selecting the first G layers, but it is not established.
  • domain assumption Users and items with similar interaction histories share similar semantic information after neighbor aggregation.
    Used in Section 3.3 to argue that heterogeneous neighbor layers form semantically relevant positive pairs. This is a standard collaborative filtering assumption but unverified for the specific datasets.
  • standard math The normalized adjacency matrix A_tilde has eigenvalues bounded by 1, with lambda_max < 1, so that propagation acts as a contraction.
    Used in Corollary 1 in Appendix A.1; for a normalized adjacency of a bipartite graph, spectral radius can be 1, not necessarily <1, so the claim of exponential decay is not guaranteed.
  • domain assumption Mutual information between adjacent layers can be computed via Gaussian differential entropy formula.
    Used in Appendix A.1 Eq (13); embeddings are not Gaussian, and the derivation of mutual information for a deterministic linear map is incorrect. This is an unsupported modeling assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NLGCL: Naturally Existing Neighbor Layers Graph Contrastive Learning for Recommendation." pith.science (2026). https://pith.science/paper/ERXMV33L

@misc{pith2026250707522,
  author       = {Pith},
  title        = {Pith review of: NLGCL: Naturally Existing Neighbor Layers Graph Contrastive Learning for Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ERXMV33L}},
  note         = {Machine review of arXiv:2507.07522}
}
read the original abstract

Graph Neural Networks (GNNs) are widely used in collaborative filtering to capture high-order user-item relationships. To address the data sparsity problem in recommendation systems, Graph Contrastive Learning (GCL) has emerged as a promising paradigm that maximizes mutual information between contrastive views. However, existing GCL methods rely on augmentation techniques that introduce semantically irrelevant noise and incur significant computational and storage costs, limiting effectiveness and efficiency. To overcome these challenges, we propose NLGCL, a novel contrastive learning framework that leverages naturally contrastive views between neighbor layers within GNNs. By treating each node and its neighbors in the next layer as positive pairs, and other nodes as negatives, NLGCL avoids augmentation-based noise while preserving semantic relevance. This paradigm eliminates costly view construction and storage, making it computationally efficient and practical for real-world scenarios. Extensive experiments on four public datasets demonstrate that NLGCL outperforms state-of-the-art baselines in effectiveness and efficiency.

Figures

Figures reproduced from arXiv: 2507.07522 by the authors.

Figure 1
Figure 1. Overview of contrastive learning. Left: traditional contrastive learning paradigm; Right: our NLGCL. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The influence of the scope of contrastive views. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Efficiency study in terms of NDCG@10. 5.4 Efficiency Study (RQ3) We provide theoretical analysis for the efficiency of our proposed NLGCL in Appendix A.2. In this section, we further validate it with an empirical study. We conduct experiments across all CL-based baselines using four datasets. In [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: We follow previous work [51] to plot feature distributions with Gaussian kernel density estimation (KDE) in R 2 (the darker the color is, the more points fall in that area.) and KDE on angles (i.e., arctan2(y, x) for each point (x,y)). 10 6 10 5 10 4 Balancing hyper-pa…
Figure 5
Figure 5. Figure 5: Performance Comparison 𝑤 .𝑟 .𝑡 . 𝜆1 and 𝜏. 1 2 3 4 Group number G 1 2 3 4 L a y e r n u m b e r L 0.0679 0.0685 0.0713 0.0692 0.0702 0.0695 0.0680 0.0688 0.0682 0.0677 Yelp 1 2 3 4 Group number G 1 2 3 4 L a y e r n u m b e r L 0.0737 0.0743 0.0755 0.0749 0.0760 0.0752…
Figure 6
Figure 6. Figure 6: Performance Comparison [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. The Best is Yet to Come: Graph Convolution in the Testing Phase for Multimodal Recommendation

    cs.IR 2025-07 conditional novelty 6.0 of 10

    A multimodal recommender that trains without graph convolution and applies it only at test time outperforms graph-trained baselines while training much faster.

Reference graph

Works this paper leans on

63 extracted references · 29 canonical work pages · cited by 1 Pith paper

  1. [1]

    Markus Bayer, Marc-André Kaufhold, and Christian Reuter. 2022. A survey on data augmentation for text classification. Comput. Surveys 55, 7 (2022), 1–39

    work page 2022

  2. [2]

    Xuheng Cai, Chao Huang, Lianghao Xia, and Xubin Ren. 2023. LightGCL: Simple Yet Effective Graph Contrastive Learning for Recommendation. In The Eleventh International Conference on Learning Representations

    2023

  3. [3]

    Lei Chen, Le Wu, Richang Hong, Kun Zhang, and Meng Wang. 2020. Revisiting graph based collaborative filtering: A linear residual graph convolutional network approach. In Proceedings of the AAAI conference on artificial intelligence , Vol. 34. 27–34

    work page 2020

  4. [4]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. In Interna- tional conference on machine learning . PMLR, 1597–1607

    2020

  5. [5]

    Tong Chen, Hongzhi Yin, Jing Long, Quoc Viet Hung Nguyen, Yang Wang, and Meng Wang. 2022. Thinking inside the box: learning hypercube representations for group recommendation. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval . 1664–1673

    work page 2022

  6. [6]

    Wen Chen, Pipei Huang, Jiaming Xu, Xin Guo, Cheng Guo, Fei Sun, Chao Li, Andreas Pfadler, Huan Zhao, and Binqiang Zhao. 2019. POG: personalized outfit generation for fashion recommendation at Alibaba iFashion. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining. 2662–2670

    work page 2019

  7. [7]

    Yen-Chi Chen. 2017. A tutorial on kernel density estimation and recent advances. Biostatistics & Epidemiology 1, 1 (2017), 161–187

    work page 2017

  8. [8]

    Zheyu Chen, Jinfeng Xu, and Haibo Hu. 2025. Don’t Lose Yourself: Boosting Multimodal Recommendation via Reducing Node-neighbor Discrepancy in Graph Convolutional Network. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 1–5

    2025

Show all 63 references
  1. [9]

    Zheyu Chen, Jinfeng Xu, Yutong Wei, and Ziyue Peng. 2025. Squeeze and Excita- tion: A Weighted Graph Contrastive Learning for Collaborative Filtering. arXiv preprint arXiv:2504.04443 (2025)

    2025 arXiv

  2. [10]

    Thomas M Cover. 1999. Elements of information theory . John Wiley & Sons

    1999

  3. [11]

    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

    2016

  4. [12]

    Wenqi Fan, Yao Ma, Qing Li, Yuan He, Eric Zhao, Jiliang Tang, and Dawei Yin

  5. [13]

    Xue Geng, Hanwang Zhang, Jingwen Bian, and Tat-Seng Chua. 2015. Learning image and user features for recommendation in social networks. In Proceedings of the IEEE international conference on computer vision . 4274–4282

    2015

  6. [14]

    Xavier Glorot and Yoshua Bengio. 2010. Understanding the difficulty of training deep feedforward neural networks. In Proceedings of the thirteenth international conference on artificial intelligence and statistics . JMLR Workshop and Conference Proceedings, 249–256

    2010

  7. [15]

    Fréderic Godin, Viktor Slavkovikj, Wesley De Neve, Benjamin Schrauwen, and Rik Van de Walle. 2013. Using topic models for twitter hashtag recommendation. In Proceedings of the 22nd international conference on world wide web . 593–596

    2013

  8. [16]

    Zhiqiang Guo, Jianjun Li, Guohui Li, Chaoyang Wang, Si Shi, and Bin Ruan. 2024. LGMRec: Local and Global Graph Learning for Multimodal Recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 38. 8454–8462

    2024

  9. [17]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. 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 ...

    2020

  10. [18]

    Mengyuan Jing, Yanmin Zhu, Tianzi Zang, and Ke Wang. 2023. Contrastive self-supervised learning in recommender systems: A survey. ACM Transactions on Information Systems 42, 2 (2023), 1–39

    2023

  11. [19]

    Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  12. [20]

    Thomas N Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In International Conference on Learning Repre- sentations

    2017

  13. [21]

    Zongwei Li, Lianghao Xia, and Chao Huang. 2024. Recdiff: Diffusion model for social recommendation. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management . 1346–1355

    2024

  14. [22]

    Zihan Lin, Changxin Tian, Yupeng Hou, and Wayne Xin Zhao. 2022. Improving graph collaborative filtering with neighborhood-enriched contrastive learning. In Proceedings of the ACM web conference 2022 . 2320–2329

    2022

  15. [23]

    Fan Liu, Zhiyong Cheng, Lei Zhu, Zan Gao, and Liqiang Nie. 2021. Interest-aware message-passing GCN for recommendation. In Proceedings of the web conference

    2021

  16. [24]

    Xiao Liu, Fanjin Zhang, Zhenyu Hou, Li Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. 2021. Self-supervised learning: Generative or contrastive. IEEE transactions on knowledge and data engineering 35, 1 (2021), 857–876

    2021

  17. [25]

    Yixin Liu, Ming Jin, Shirui Pan, Chuan Zhou, Yu Zheng, Feng Xia, and S Yu Philip

  18. [26]

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

    2018 arXiv

  19. [27]

    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 ACM Web Conference 2023 . 1097–1108

    2023

  20. [28]

    Sylvestre-Alvise Rebuffi, Sven Gowal, Dan Andrei Calian, Florian Stimberg, Olivia Wiles, and Timothy A Mann. 2021. Data augmentation can improve robustness. Advances in Neural Information Processing Systems 34 (2021), 29935–29948

    2021

  21. [29]

    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

  22. [31]

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

  23. [32]

    Aravind Sankar, Yanhong Wu, Yuhang Wu, Wei Zhang, Hao Yang, and Hari Sundaram. 2020. Groupim: A mutual information maximization framework for neural group recommendation. InProceedings of the 43rd International ACM SIGIR conference on research and development in Information Re...

    2020

  24. [33]

    Brent Smith and Greg Linden. 2017. Two decades of recommender systems at Amazon. com. Ieee internet computing 21, 3 (2017), 12–18

    2017

  25. [34]

    Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-SNE. Journal of machine learning research 9, 11 (2008)

    2008

  26. [35]

    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

  27. [36]

    Tongzhou Wang and Phillip Isola. 2020. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning . PMLR, 9929–9939

    2020

  28. [37]

    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

  29. [38]

    Wei Wei, Chao Huang, Lianghao Xia, and Chuxu Zhang. 2023. Multi-Modal Self-Supervised Learning for Recommendation. In Proceedings of the ACM Web Conference 2023. 790–800

    2023

  30. [39]

    Felix Wu, Amauri Souza, Tianyi Zhang, Christopher Fifty, Tao Yu, and Kilian Weinberger. 2019. Simplifying graph convolutional networks. In International conference on machine learning . PMLR, 6861–6871

    2019

  31. [40]

    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

  32. [41]

    Xixi Wu, Yun Xiong, Yao Zhang, Yizhu Jiao, Jiawei Zhang, Yangyong Zhu, and Philip S Yu. 2023. Consrec: Learning consensus behind interactions for group recommendation. In Proceedings of the ACM Web Conference 2023 . 240–250

    2023

  33. [42]

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

  34. [43]

    Jinfeng Xu, Zheyu Chen, Jinze Li, Shuo Yang, Hewei Wang, Yijie Li, Mengran Li, Puzhen Wu, and Edith CH Ngai. 2025. MDVT: Enhancing Multimodal Rec- ommendation with Model-Agnostic Multimodal-Driven Virtual Triplets. arXiv preprint arXiv:2505.16665 (2025)

    2025 arXiv

  35. [44]

    Jinfeng Xu, Zheyu Chen, Jinze Li, Shuo Yang, Hewei Wang, and Edith CH Ngai

  36. [45]

    Jinfeng Xu, Zheyu Chen, Jinze Li, Shuo Yang, Wei Wang, Xiping Hu, and Edith C-H Ngai. 2024. FourierKAN-GCF: Fourier Kolmogorov-Arnold Network–An Effective and Efficient Feature Transformation for Graph Collaborative Filtering. arXiv preprint arXiv:2406.01034 (2024)

    2024 arXiv

  37. [46]

    Jinfeng Xu, Zheyu Chen, Zixiao Ma, Jiyi Liu, and Edith CH Ngai. 2024. Im- proving Consumer Experience With Pre-Purify Temporal-Decay Memory-Based Collaborative Filtering Recommendation for Graduate School Application. IEEE Transactions on Consumer Electronics (2024)

    2024

  38. [47]

    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

  39. [48]

    Jinfeng Xu, Zheyu Chen, Shuo Yang, Jinze Li, Hewei Wang, and Edith CH Ngai

  40. [49]

    Jinfeng Xu, Zheyu Chen, Shuo Yang, Jinze Li, Wei Wang, Xiping Hu, Steven Hoi, and Edith Ngai. 2025. A Survey on Multimodal Recommender Systems: Recent Advances and Future Directions. arXiv preprint arXiv:2502.15711 (2025)

    2025 arXiv

  41. [50]

    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

  42. [51]

    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. In Proceedings of the 45th international ACM SIGIR conference on research and development in info...

    2022

  43. [52]

    Guanghu Yuan, Fajie Yuan, Yudong Li, Beibei Kong, Shujie Li, Lei Chen, Min Yang, Chenyun Yu, Bo Hu, Zang Li, et al . 2022. Tenrec: A large-scale multipurpose benchmark dataset for recommender systems. Advances in Neural Information Processing Systems 35 (2022), 11480–11493

    2022

  44. [53]

    Jinfeng Xu, Zheyu Chen, Wei Wang, Xiping Hu, Sang-Wook Kim, and Edith CH Ngai. 2025. COHESION: Composite Graph Convolutional Network with Dual- Stage Fusion for Multimodal Recommendation. arXiv preprint arXiv:2504.04452 (2025)

    2025 arXiv

  45. [54]

    Yi Zhang, Lei Sang, and Yiwen Zhang. 2024. Exploring the Individuality and Collectivity of Intents behind Interactions for Graph Collaborative Filtering. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval . 1253–1262

    2024

  46. [55]

    Wayne Xin Zhao, Yupeng Hou, Xingyu Pan, Chen Yang, Zeyu Zhang, Zihan Lin, Jingsen Zhang, Shuqing Bian, Jiakai Tang, Wenqi Sun, et al. 2022. RecBole 2.0: towards a more up-to-date recommendation library. In Proceedings of the 31st ACM International Conference on Information & K...

    2022

  47. [56]

    Xin Zhou, Donghui Lin, Yong Liu, and Chunyan Miao. 2023. Layer-refined graph convolutional networks for recommendation. In 2023 IEEE 39th International Conference on Data Engineering (ICDE) . IEEE, 1247–1259

    2023

  48. [57]

    Xin Zhou and Zhiqi Shen. 2023. A tale of two graphs: Freezing and denoising graph structures for multimodal recommendation. In Proceedings of the 31st ACM International Conference on Multimedia . 935–943

    2023

  49. [60]

    Dan Zhang, Yangliao Geng, Wenwen Gong, Zhongang Qi, Zhiyu Chen, Xing Tang, Ying Shan, Yuxiao Dong, and Jie Tang. 2024. RecDCL: Dual Contrastive Learning for Recommendation. In Proceedings of the ACM on Web Conference

    2024

  50. [2009]

    In Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence

    BPR: Bayesian personalized ranking from implicit feedback. In Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence . 452–461

  51. [2012]

    arXiv preprint arXiv:1205.2618 (2012)

    BPR: Bayesian personalized ranking from implicit feedback. arXiv preprint arXiv:1205.2618 (2012)

    2012 arXiv

  52. [2019]

    In The world wide web conference

    Graph neural networks for social recommendation. In The world wide web conference. 417–426

  53. [2022]

    IEEE transactions on knowledge and data engineering 35, 6 (2022), 5879–5900

    Graph self-supervised learning: A survey. IEEE transactions on knowledge and data engineering 35, 6 (2022), 5879–5900

    2022

  54. [2024]

    In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management

    AlignGroup: Learning and Aligning Group Consensus with Member Pref- erences for Group Recommendation. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management . 2682–2691

  55. [2025]

    In Proceedings of the AAAI Conference on Artificial Intelligence , Vol

    Mentor: multi-level self-supervised learning for multimodal recommen- dation. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 39. 12908–12917. NLGCL RecSys ’25, September 22–26, 2025, Prague, Czech Republic

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.