Pith. sign in

REVIEW 5 major objections 6 minor 51 references

Matrix Factorization with Dynamic Multi-view Clustering for Recommender System

T0 review · 5 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read The paper claims that replacing free latent vectors with weighted, pruned cluster centers across views makes matrix factorization both more accurate and more interpretable.

desk verdict MFDMC is a worthwhile synthesis of multi-view clustering and matrix factorization with plausible RMSE gains, but a false normalization claim in Eq. 7 undercuts the stated balancing mechanism and needs fixing. read the letter →

arxiv 2504.14565 v1 pith:LN52ONBU submitted 2025-04-20 cs.IR cs.SI

Shangde Gao , Ke Liu , Yichao Fu , Hongxia Xu , Jian Wu This is my paper · ORCID
classification cs.IRcs.SI
keywords matrixfactorizationmulti-viewclusteringrecommendersystemsrepresentationlearningdynamicinterpretabilitylatentspaceutilizationRMSE
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 claims that standard matrix factorization leaves much of its latent space unused, and that replacing free user and item vectors with weighted combinations of cluster centers organized into several views, learned end-to-end with clustering losses and dynamic pruning, remedies that waste. On six recommendation datasets it reports lower RMSE with a latent dimension of 16 than baselines achieve at dimension 60, including an improvement of about 0.025 RMSE over the FunkMF baseline on MovieLens-100k. It further reports that clusters take on interpretable meanings, such as movie genres, so a recommendation can be traced back to the clusters that dominate a user's representation. The same dynamic multi-view clustering module, attached to CNN features, is reported to transfer to image classification on a synthetic shape-and-color dataset. At stake is whether recommenders can be smaller, faster, and explainable without losing accuracy.

What carries the argument

The central object is the multi-view weighted-cluster representation: each user or item latent vector is formed by concatenating, over $v$ views, the weighted sum of $t$ cluster centers, with the user form written as $\mathbf{p}_i = \bigoplus_{j=0}^{v} \sum_{i'=0}^{t} \mathbf{c}^{\mathrm{user}}_{i',j}\, w^{\mathrm{user}}_{i',j}$ and an analogous formula for items. Cluster centers are learned with a spread loss and a proximity MSE, while the weights are softmax-normalized and then reweighted by the mapping function of Eq. 7 so that entropy losses are claimed to be comparable across views. A pruning step, run every $I_p$ epochs, removes centers whose mean weight falls below a threshold, so the number of centers per view changes during training. This construction carries the argument: it is the mechanism intended to use the representation space fully, produce interpretable clusters, and keep optimization balanced across views.

What would settle it

Substitute uniform weights into Eq. 7 and Eq. 8 for two views with different numbers of centers: the resulting entropy losses are $t/e$ for $t$ centers, so the claimed invariance fails; a reader can verify this with a one-line calculation, and it directly undermines the stated purpose of the mapping function.

Watch

Extended reading notes

Core claim

MFDMC's central discovery is that a matrix-factorization model can be built so that every user and item vector is a concatenation of weighted sums of cluster centers across several views, and that this construction, optimized with a spread loss that keeps centers apart, a proximity loss that binds points to centers, an entropy loss on the weights, and a pruning rule that deletes low-weight centers, makes better use of the representation space than unconstrained latent vectors. On the paper's six datasets, MFDMC with latent dimension 16 reports RMSE below FunkMF(60), BiasedMF(60), PMF(60), SVD, and Glocal-K; on MovieLens-100k the gap to FunkMF is about 0.025 RMSE. The authors also report that the learned clusters correspond to semantic categories, and that the same module transfers to image classification.

Load-bearing premise

The load-bearing premise is that the reweighting formula in Eq. 7 makes the worst-case entropy loss independent of the number of clusters in a view; substituting uniform weights contradicts this because the loss grows with the cluster count.

Editorial extensions

If this is right

  • At latent dimension 16, MFDMC reports lower RMSE than baselines run at dimension 60, so recommender models could run with far fewer parameters for the same accuracy.
  • Because clustering happens inside the same training loop, the separate clustering stage of two-stage pipelines disappears, reducing computational cost at web scale.
  • User and item embeddings become sums of view-specific cluster centers, so a recommendation can be explained by which clusters, for example genre clusters, dominate the user's vector.
  • The same dynamic clustering module, attached to CNN features, transfers to image classification, suggesting the mechanism is not specific to ratings data.
  • Dynamic pruning means the number of clusters is not fixed in advance; views can shrink during training, adapting to the data rather than requiring a grid search over cluster counts.

Reading between the lines

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

  • Beyond the paper: if the Eq.-7 balancing property fails, the reported gains may be carried by the spread and proximity losses instead, which a simple ablation that removes only the mapping function could isolate.
  • Beyond the paper: dynamic pruning currently only removes weak centers; adding a splitting operation would let the model track emerging item categories or user interests in streaming data.
  • Beyond the paper: the interpretability claim implies a testable behavior, namely that users whose vectors are dominated by a high-rated animation cluster should receive animation-heavy recommendations, which could be checked on held-out interactions.
  • Beyond the paper: the same weighted-cluster representation could be applied to other embedding-based models, such as graph embeddings, where a node's multiple roles could be captured by separate views.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes MFDMC, a matrix-factorization recommender model that learns user and item representations as weighted combinations of cluster centers across multiple views, with dynamic pruning of clusters and an entropy-based regularizer on the weights. The authors report RMSE improvements over FunkMF, BiasedMF, PMF, SVD, and Glocal-K on six datasets, an ablation study on MovieLens-1M, and a small synthetic computer-vision experiment to illustrate interpretability and transferability. The core claimed contribution is that dynamic multi-view clustering makes better use of the latent representation space, yielding more accurate and more interpretable recommendations at low dimensionality.

Significance. If the method and experiments were fully valid, the idea of coupling matrix factorization with dynamic multi-view clustering would be a useful and timely contribution to representation learning for recommender systems. The paper's strengths include evaluation on multiple public datasets, an ablation study, a visualization-based interpretability analysis, and an extension to a downstream image-classification task. However, the paper as written contains a demonstrable arithmetic error in the central balancing mechanism, incomplete formal specification of the model, and an experimental section without variance estimates or modern baselines. These issues currently prevent the claims from being accepted at face value.

major comments (5)
  1. [Section III.C, Eqs. (7)-(8)] The claimed invariance of the uniform-distribution entropy loss to the number of cluster centers t is arithmetically false. Substituting w' = 1/t into Eq. (7) gives w'' = 1/e for every center, so Eq. (8) yields loss = t/e, which grows linearly with t. A view with 10 centers has 10/3 times the loss of a view with 3 centers under the same uniform distribution. Since this invariance is the stated motivation for the mapping function and the basis for balanced multi-view optimization, the explanatory mechanism is unsupported. Please correct the mapping, add a renormalization step, or revise the claim.
  2. [Section III.A, Eq. (2)] The notation in Eq. (2) is internally inconsistent and prevents reproduction. The user index i is reused as the cluster-center index, the concatenation operator is not defined, and the summation bounds are unclear. Similarly, Eq. (8) sums w^{user''}_{i,j} over i = 0 to m, but Table I defines W as a v x e matrix without a per-user dimension. Please rewrite the model with distinct indices for users, views, and centers, and state the exact tensor shapes of all weight variables.
  3. [Section IV.C, Tables III-IV] The central claim that MFDMC consistently outperforms competitors is supported only by single-point RMSE values with no error bars, repeated runs, or significance tests. The baselines are also limited to classic matrix-factorization methods; no recent deep MF, multi-view clustering, or graph-based recommender is included, despite the related-work section positioning the paper against such methods. Please add repeated-run statistics with variance reporting and compare against at least one or two modern strong baselines.
  4. [Sections I and IV] The abstract and introduction make strong claims about substantial reductions in time and computational demands and about scalability to web-scale data, but the experiments contain no runtime, memory, or convergence measurements anywhere. Tables III and IV report only RMSE. Please provide an empirical complexity evaluation or a formal complexity analysis, or temper the computational-efficiency claims to what the experiments actually demonstrate.
  5. [Section IV.B and Table V] The text states that "centers in the same location in the view can be shared" between users and items, but Eqs. (2)-(5) define separate C^{user} and C^{item} variables and do not reflect sharing. If shared centers are used in the reported experiments, the objective must be amended to show how sharing is realized; otherwise the distinction between "Not Share" and "Share" in Table V is unclear. Please specify which equations govern the shared-center configuration.
minor comments (6)
  1. [Section III.B, Eq. (4)] The summation notation in Eq. (4) is malformed: "N,k∈Si,jX k=0" is not a well-formed expression. Please rewrite it with explicit index ranges.
  2. [Table I] The notation "e,v ∈ Z" appears to be a typo for "t,v ∈ Z", and the constraint b = d/v requires that v divides d; please state this divisibility assumption explicitly.
  3. [Section III.C] The example comparing view1 with 3 centers and view2 with 10 centers appears to claim the opposite of the intended effect: under a uniform distribution, the total entropy loss should be smaller for fewer centers, not larger. Please clarify whether the text is describing per-center or total loss.
  4. [Section III.C, Eq. (6)] Please specify the axis along which softmax is applied: softmax over the centers within each view, or over all weights jointly.
  5. [Algorithm 1] The instruction "W ← cluster-wise mean of user/item in W" is ambiguous; please define precisely what is averaged and over which subset of users or items.
  6. [Throughout] There are several typos, including "FunckMF" in the Introduction, "RSME" in Eq. (19), and "MovieLen" instead of "MovieLens" in Section IV.A. Also, references [39] and [42] appear to be the same paper (Explainable Matrix Factorization) and should be merged or disambiguated.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MFDMC's performance claims are benchmarked against external baselines on public datasets, and the cited self-works are not load-bearing for the central result.

full rationale

The paper's central derivation is self-contained in the sense required by the circularity test: MFDMC's representation is defined directly by Eq. 2 (weighted sums of cluster centers), the losses are specified in Eqs. 3-10, and the recommendation quality is then measured with RMSE against independent baselines (FunkMF, BiasedMF, PMF, SVD, Glocal-K) on public datasets. No fitted parameter is renamed as a prediction; hyperparameters are tuned with Optuna on a validation split and test RMSE is reported. The self-citations ([14], [15], [24], [26], [27], [38]) appear in the related-work discussion and are not used to justify the architecture or the empirical gains. The one notable defect is Section III-C's claim that the Eq. 7 mapping makes uniform-distribution entropy loss independent of the number of centers t; substituting w' = 1/t gives w'' = 1/e for each center and loss t/e in Eq. 8, so the stated balancing property is arithmetically false. That is a correctness/soundness error in a design rationale, not a circular step: the paper does not define the mapping in terms of the loss it is supposed to produce, and the empirical comparison does not reduce to the mapping's own outputs.

Assumptions & free parameters 8 free parameters · 5 assumptions · 1 invented entities

The central claim depends on several per-dataset hyperparameters (rho, eta, gamma, v, d, t, psi, Id) and on the unverified claim that the Eq. 7 mapping balances losses across views. These choices are fitted to data rather than derived from first principles.

free parameters (8)
  • rho (spread loss margin) = not reported per dataset
    Controls the minimum required distance between cluster centers; tuned by Optuna.
  • eta (loss1 weight) = not reported
    Balances cluster-center spread and proximity losses against the rating loss; grows with epochs and is tuned by Optuna.
  • gamma (loss2 weight) = not reported
    Balances the weight entropy loss against the rating loss; tuned by Optuna.
  • psi (pruning threshold) = 1/t
    Cluster centers whose mean weight falls below 1/t are pruned; this value is chosen by hand.
  • number of views v = 2, 4, 6, 8, or 10 per dataset
    Splits the latent space into views; selected by Optuna and varied in the ablation.
  • initial cluster count t per view = 10
    The number of cluster centers before pruning; fixed across experiments.
  • latent dimension d = 12 or 16
    The dimension of user and item vectors; chosen manually and tested in the ablation.
  • Id (pruning start epoch) = 40
    The first 40 epochs keep all cluster centers fixed; the threshold value is stated in Section IV-B.
assumptions (5)
  • domain assumption Ratings are approximated by inner products of user and item latent vectors.
    Invoked as loss3 in Eq. 10 and throughout the experiments; this is the standard MF assumption.
  • domain assumption Euclidean distance is an appropriate metric for cluster spread and proximity.
    Used in Eq. 3, Eq. 4 and Eq. 17; no justification for choosing Euclidean over other metrics.
  • ad hoc to paper The mapping function in Eq. 7 makes the worst-case entropy loss invariant to the number of cluster centers t per view.
    The paper states this in Section III-C, but substituting uniform weights into Eq. 7 and Eq. 8 gives loss = t/e, which varies with t; the axiom is contradicted by the equations.
  • domain assumption Softmax normalization of cluster weights is appropriate for interpretability and optimization.
    Applied in Eq. 6 and Eq. 14; no alternatives are discussed.
  • ad hoc to paper Pruning low-weight clusters does not degrade representation quality.
    Algorithm 1 removes clusters with mean weight below psi; the effect on downstream RMSE is not analyzed.
invented entities (1)
  • Shared cluster centers between users and items
    purpose: Reduce parameter count by using the same center vectors in corresponding view positions for both users and items.
    Mentioned in Section IV-B; this is a modeling choice with no external support.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Matrix Factorization with Dynamic Multi-view Clustering for Recommender System." pith.science (2026). https://pith.science/paper/LN52ONBU

@misc{pith2026250414565,
  author       = {Pith},
  title        = {Pith review of: Matrix Factorization with Dynamic Multi-view Clustering for Recommender System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LN52ONBU}},
  note         = {Machine review of arXiv:2504.14565}
}
read the original abstract

Matrix factorization (MF), a cornerstone of recommender systems, decomposes user-item interaction matrices into latent representations. Traditional MF approaches, however, employ a two-stage, non-end-to-end paradigm, sequentially performing recommendation and clustering, resulting in prohibitive computational costs for large-scale applications like e-commerce and IoT, where billions of users interact with trillions of items. To address this, we propose Matrix Factorization with Dynamic Multi-view Clustering (MFDMC), a unified framework that balances efficient end-to-end training with comprehensive utilization of web-scale data and enhances interpretability. MFDMC leverages dynamic multi-view clustering to learn user and item representations, adaptively pruning poorly formed clusters. Each entity's representation is modeled as a weighted projection of robust clusters, capturing its diverse roles across views. This design maximizes representation space utilization, improves interpretability, and ensures resilience for downstream tasks. Extensive experiments demonstrate MFDMC's superior performance in recommender systems and other representation learning domains, such as computer vision, highlighting its scalability and versatility.

Figures

Figures reproduced from arXiv: 2504.14565 by the authors.

Figure 1
Figure 1. Schematic Illustration of the Multi-View Feature-Weighted Dynamic [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Mapping function of Eq. 7 for the weight. Using the function in the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the MFDMC for computer vision tasks. The features [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Multi-view Clustering results. The four rows show the clustering centers of users/items, and the clustering results of users/items respectively. Each [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: In the color view of Fig. 5(a), the green, red and [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 5
Figure 5. Figure 5: The t-sne results of experiments on synthetic toy dataset. (a), (b), and (c) show the results of shape, color view, and the entire embedding respectively. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Illustration of interpreting a representation. Red circles depict the [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 38 canonical work pages

  1. [1]

    Self-supervised learning for recommender systems: A survey,

    J. Yu, H. Yin, X. Xia, T. Chen, J. Li, and Z. Huang, “Self-supervised learning for recommender systems: A survey,” TKDE, 2023

    work page 2023

  2. [2]

    A survey of recommendation systems: recommendation models, techniques, and application fields,

    H. Ko, S. Lee, Y . Park, and A. Choi, “A survey of recommendation systems: recommendation models, techniques, and application fields,” Electronics, vol. 11, no. 1, p. 141, 2022

    2022

  3. [3]

    Recommender systems in the era of large language models (llms),

    Z. Zhao, W. Fan, J. Li, Y . Liu, X. Mei, Y . Wang, Z. Wen, F. Wang, X. Zhao, J. Tang et al. , “Recommender systems in the era of large language models (llms),” IEEE Transactions on Knowledge and Data Engineering, 2024

    2024

  4. [4]

    Matrix factorization in recommender systems: algo- rithms, applications, and peculiar challenges,

    F. O. Isinkaye, “Matrix factorization in recommender systems: algo- rithms, applications, and peculiar challenges,” IETE Journal of Research, vol. 69, no. 9, pp. 6087–6100, 2023

    work page 2023

  5. [5]

    Comprehensive Evaluation of Matrix Factorization Models for Collaborative Filtering Recommender Systems

    J. Bobadilla, J. Due ˜nas-Ler´ın, F. Ortega, and A. Gutierrez, “Comprehen- sive evaluation of matrix factorization models for collaborative filtering recommender systems,” arXiv preprint arXiv:2410.17644 , 2024

    work page Pith review arXiv 2024

  6. [6]

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

    A. Xiang, B. Huang, X. Guo, H. Yang, and T. Zheng, “A neural matrix decomposition recommender system model based on the multimodal large language model,” arXiv preprint arXiv:2407.08942 , 2024

    arXiv 2024

  7. [7]

    Surprise: A python library for recommender systems,

    N. Hug, “Surprise: A python library for recommender systems,” Journal of Open Source Software , vol. 5, no. 52, p. 2174, 2020. [Online]. Available: https://doi.org/10.21105/joss.02174

    doi:10.21105/joss.02174 2020

  8. [8]

    Matrix factorization techniques for recommender systems,

    Y . Koren, R. Bell, and C. V olinsky, “Matrix factorization techniques for recommender systems,” Computer, vol. 42, no. 8, pp. 30–37, 2009

    2009

Show all 51 references
  1. [9]

    Tt-svd: An efficient sparse decision-making model with two- way trust recommendation in the ai-enabled iot systems,

    G. Xu, Y . Zhao, L. Jiao, M. Feng, Z. Ji, E. Panaousis, S. Chen, and X. Zheng, “Tt-svd: An efficient sparse decision-making model with two- way trust recommendation in the ai-enabled iot systems,” IEEE Internet of Things Journal , vol. 8, no. 12, pp. 9559–9567, 2020

    2020

  2. [10]

    Self- adaptive deep asymmetric network for imbalanced recommendation,

    Y . Zhu, Y . Geng, Y . Li, J. Qiang, Y . Yuan, and X. Wu, “Self- adaptive deep asymmetric network for imbalanced recommendation,” IEEE Transactions on Emerging Topics in Computational Intelligence , 2023

    2023

  3. [11]

    When federated recommendation meets cold-start problem: Separating item attributes and user interactions,

    C. Zhang, G. Long, T. Zhou, Z. Zhang, P. Yan, and B. Yang, “When federated recommendation meets cold-start problem: Separating item attributes and user interactions,” in Proceedings of the ACM on Web Conference 2024, 2024, pp. 3632–3642

    2024

  4. [12]

    Personalized news recommen- dation: Methods and challenges,

    C. Wu, F. Wu, Y . Huang, and X. Xie, “Personalized news recommen- dation: Methods and challenges,” TOIS, vol. 41, no. 1, pp. 1–50, 2023

    2023

  5. [13]

    Paper recommendation using specter with low-rank and sparse matrix factorization

    P. Guo, G. Zhou, J. Lu, Z. Li, and T. Zhu, “Paper recommendation using specter with low-rank and sparse matrix factorization.” KSII Transactions on Internet & Information Systems , vol. 18, no. 5, 2024

    2024

  6. [14]

    Contrastive knowledge amalgama- tion for unsupervised image classification,

    S. Gao, Y . Fu, K. Liu, and Y . Han, “Contrastive knowledge amalgama- tion for unsupervised image classification,” 2023

    2023

  7. [15]

    Collaborative knowledge amalgamation: Preserving discriminability and transferability in unsupervised learning,

    S. Gao, Y . Fu, K. Liu, W. Gao, H. Xu, J. Wu, and Y . Han, “Collaborative knowledge amalgamation: Preserving discriminability and transferability in unsupervised learning,” Information Sciences , vol. 669, p. 120564, 2024

    2024

  8. [16]

    Multi-view clustering via deep matrix factorization,

    H. Zhao, Z. Ding, and Y . Fu, “Multi-view clustering via deep matrix factorization,” in Proceedings of the AAAI conference on artificial intelligence, vol. 31, no. 1, 2017

    2017

  9. [17]

    A survey of graph neural network based recommendation in social networks,

    X. Li, L. Sun, M. Ling, and Y . Peng, “A survey of graph neural network based recommendation in social networks,” Neurocomputing, vol. 549, p. 126441, 2023

    2023

  10. [18]

    Dis- tribution knowledge embedding for graph pooling,

    K. Chen, J. Song, S. Liu, N. Yu, Z. Feng, G. Han, and M. Song, “Dis- tribution knowledge embedding for graph pooling,” IEEE Transactions on Knowledge and Data Engineering , 2023

    2023

  11. [19]

    Improving expressivity of gnns with subgraph-specific factor embedded normalization,

    K. Chen, S. Liu, T. Zhu, J. Qiao, Y . Su et al., “Improving expressivity of gnns with subgraph-specific factor embedded normalization,” in ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2023

    2023

  12. [20]

    Self-weighted contrastive fusion for deep multi-view clustering,

    S. Wu, Y . Zheng, Y . Ren, J. He, X. Pu, S. Huang, Z. Hao, and L. He, “Self-weighted contrastive fusion for deep multi-view clustering,” IEEE Transactions on Multimedia , 2024

    2024

  13. [21]

    A generalized deep learning clustering algorithm based on non-negative matrix factorization,

    D. Wang, T. Li, P. Deng, F. Zhang, W. Huang, P. Zhang, and J. Liu, “A generalized deep learning clustering algorithm based on non-negative matrix factorization,” KDD, vol. 17, no. 7, pp. 1–20, 2023

    2023

  14. [22]

    Deepfm: a factorization- machine based neural network for ctr prediction,

    H. Guo, R. Tang, Y . Ye, Z. Li, and X. He, “Deepfm: a factorization- machine based neural network for ctr prediction,” arXiv preprint arXiv:1703.04247, 2017

    2017 arXiv

  15. [23]

    Dc- cnmf: Deep complementary and consensus non-negative matrix factor- ization for multi-view clustering,

    S. Gunawardena, K. Luong, T. Balasubramaniam, and R. Nayak, “Dc- cnmf: Deep complementary and consensus non-negative matrix factor- ization for multi-view clustering,” Knowledge-Based Systems, vol. 285, p. 111330, 2024

    2024

  16. [24]

    Interest-oriented universal user representation via contrastive learning,

    Q. Sun, J. Gu, B. Yang, X. Xu, R. Xu, S. Gao, H. Liu, and H. Xu, “Interest-oriented universal user representation via contrastive learning,” arXiv preprint arXiv:2109.08865 , 2021

    2021 arXiv

  17. [25]

    An attention-based framework for multi-view clustering on grassmann manifold,

    D. Wu, X. Dong, F. Nie, R. Wang, and X. Li, “An attention-based framework for multi-view clustering on grassmann manifold,” Pattern Recognition, vol. 128, p. 108610, 2022

    2022

  18. [26]

    Learn- ing interest-oriented universal user representation via self-supervision,

    Q. Sun, J. Gu, X. Xu, R. Xu, K. Liu, B. Yang, H. Liu, and H. Xu, “Learn- ing interest-oriented universal user representation via self-supervision,” in Proceedings of the 30th ACM International Conference on Multime- dia, 2022, pp. 7270–7278

    2022

  19. [27]

    Empow- ering general-purpose user representation with full-life cycle behavior modeling,

    B. Yang, J. Gu, K. Liu, X. Xu, R. Xu, Q. Sun, and H. Liu, “Empow- ering general-purpose user representation with full-life cycle behavior modeling,” in Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2023, pp. 2908–2917

    2023

  20. [28]

    Recommender systems survey,

    J. Bobadilla, F. Ortega, A. Hernando, and A. Guti ´errez, “Recommender systems survey,” KBS, vol. 46, pp. 109–132, 2013. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0950705113001044

    2013

  21. [29]

    Dnn-mf: Deep neural network matrix factorization approach for filtering information in multi-criteria recommender systems,

    B. B. Sinha and R. Dhanalakshmi, “Dnn-mf: Deep neural network matrix factorization approach for filtering information in multi-criteria recommender systems,” Neural Computing and Applications , vol. 34, no. 13, pp. 10 807–10 821, 2022

    2022

  22. [30]

    Netflix update: Try this at home,

    Funk, Simon, “Netflix update: Try this at home,” https://sifter.org/ ∼simon/journal/20061211.html, 2006, accessed: 2021-08-20

    2006

  23. [31]

    Probabilistic matrix factorization,

    A. Mnih and R. R. Salakhutdinov, “Probabilistic matrix factorization,” in NeurIPS, 2008, pp. 1257–1264

    2008

  24. [32]

    Lightfr: Lightweight federated recommendation with privacy-preserving matrix factorization,

    H. Zhang, F. Luo, J. Wu, X. He, and Y . Li, “Lightfr: Lightweight federated recommendation with privacy-preserving matrix factorization,” TIS, vol. 41, no. 4, pp. 1–28, 2023

    2023

  25. [33]

    Robust multi-view clustering with noisy correspondence,

    Y . Sun, Y . Qin, Y . Li, D. Peng, X. Peng, and P. Hu, “Robust multi-view clustering with noisy correspondence,”IEEE Transactions on Knowledge and Data Engineering , 2024

    2024

  26. [34]

    Prototype matching learning for incomplete multi-view clustering,

    H. Yuan, Y . Sun, F. Zhou, J. Wen, S. Yuan, X. You, and Z. Ren, “Prototype matching learning for incomplete multi-view clustering,” IEEE Transactions on Image Processing , 2025

    2025

  27. [35]

    A survey of visual transformers,

    Y . Liu, Y . Zhang, Y . Wang, F. Hou, J. Yuan, J. Tian, Y . Zhang, Z. Shi, J. Fan, and Z. He, “A survey of visual transformers,” TNNLS, pp. 1–21, 2023

    2023

  28. [36]

    Better integrating vision and semantics for improving few-shot classification,

    Z. Li and Y . Wang, “Better integrating vision and semantics for improving few-shot classification,” in Proceedings of the 31st ACM international conference on multimedia , 2023, pp. 4737–4746

    2023

  29. [37]

    Fewvs: A vision-semantics integration framework for few-shot image classification,

    Z. Li, Y . Wang, and K. Li, “Fewvs: A vision-semantics integration framework for few-shot image classification,” inProceedings of the 32nd ACM International Conference on Multimedia , 2024, pp. 1341–1350

    2024

  30. [38]

    Everyone’s preference changes differently: A weighted multi-interest model for retrieval,

    H. Shi, Y . Gu, Y . Zhou, B. Zhao, S. Gao, and J. Zhao, “Everyone’s preference changes differently: A weighted multi-interest model for retrieval,” in ICML. PMLR, 2023, pp. 31 228–31 242

    2023

  31. [40]

    Learning to rank features for recommendation over multiple categories,

    X. Chen, Z. Qin, Y . Zhang, and T. Xu, “Learning to rank features for recommendation over multiple categories,” inSIGIR, 2016, pp. 305–314

    2016

  32. [41]

    Aspect based recommendations: Recommending items with the most valuable aspects based on user reviews,

    K. Bauman, B. Liu, and A. Tuzhilin, “Aspect based recommendations: Recommending items with the most valuable aspects based on user reviews,” in SIGKDD, 2017, pp. 717–725

    2017

  33. [42]

    Explainable matrix factorization for collaborative filtering,

    B. Abdollahi and O. Nasraoui, “Explainable matrix factorization for collaborative filtering,” in WWW, 2016, pp. 5–6

    2016

  34. [43]

    Visual interpretability of image-based classification models by generative latent space dis- entanglement applied to in vitro fertilization,

    O. Rotem, T. Schwartz, R. Maor, Y . Tauber, M. T. Shapiro, M. Meseguer, D. Gilboa, D. S. Seidman, and A. Zaritsky, “Visual interpretability of image-based classification models by generative latent space dis- entanglement applied to in vitro fertilization,” Nature communicatio...

    2024

  35. [44]

    Visual interpretability of bioimaging deep learning models,

    O. Rotem and A. Zaritsky, “Visual interpretability of bioimaging deep learning models,” nature methods, vol. 21, no. 8, pp. 1394–1397, 2024

    2024

  36. [45]

    The movielens datasets: History and context,

    F. M. Harper and J. A. Konstan, “The movielens datasets: History and context,” vol. 5, no. 4, Dec. 2015. [Online]. Available: https://doi.org/10.1145/2827872

    2015 doi

  37. [46]

    Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering,

    R. He and J. McAuley, “Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering,” in WWW, 2016, pp. 507–517

    2016

  38. [47]

    Improving recommendation lists through topic diversification,

    C.-N. Ziegler, S. M. McNee, J. A. Konstan, and G. Lausen, “Improving recommendation lists through topic diversification,” in WWW, 2005, pp. 22–32

    2005

  39. [48]

    Eigentaste: A constant time collaborative filtering algorithm,

    K. Goldberg, T. Roeder, D. Gupta, and C. Perkins, “Eigentaste: A constant time collaborative filtering algorithm,” information retrieval , vol. 4, no. 2, pp. 133–151, 2001

    2001

  40. [49]

    Optuna: A next- generation hyperparameter optimization framework,

    T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama, “Optuna: A next- generation hyperparameter optimization framework,” in Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining , 2019, pp. 2623–2631

    2019

  41. [50]

    Neural network matrix factorization,

    G. K. Dziugaite and D. M. Roy, “Neural network matrix factorization,” arXiv preprint arXiv:1511.06443 , 2015

    2015 arXiv

  42. [51]

    A comprehensive survey of evaluation techniques for recommendation systems,

    A. Jadon and A. Patil, “A comprehensive survey of evaluation techniques for recommendation systems,” 2023

    2023

  43. [52]

    Glocal-k: Global and local kernels for recommender systems,

    S. C. Han, T. Lim, S. Long, B. Burgstaller, and J. Poon, “Glocal-k: Global and local kernels for recommender systems,” in Proceedings of the 30th ACM International Conference on Information & Knowledge Management, 2021, pp. 3063–3067

    2021

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.