Pith. sign in

REVIEW 4 major objections 6 minor 51 references

DiffCL: A Diffusion-Based Contrastive Learning Framework with Semantic Alignment for Multimodal Recommendations

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read DiffCL claims that diffusion-generated contrastive views, ID-guided alignment, and an item-item graph together achieve state-of-the-art multimodal recommendation on the Baby, Video, and Sports datasets.

desk verdict DiffCL reports consistent multimodal recommendation gains, but its core diffusion module is under-specified to the point that the method as written is not a real diffusion model—missing a denoising training objective entirely. read the letter →

arxiv 2501.01066 v1 pith:M4XGPNJQ submitted 2025-01-02 cs.MM

Qiya Song , Jiajun Hu , Lin Xiao , Bin Sun , Xieping Gao , Shutao Li This is my paper · ORCID
classification cs.MM
keywords multimodalrecommendationdiffusionmodelgraphcontrastivelearningsemanticalignmentitem-itemself-supervisedcross-modalfusionrecommendersystems
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 DiffCL, a multimodal recommendation framework that changes how contrastive views are built in graph-based recommender systems. Instead of edge dropout or random noise, a diffusion forward process adds Gaussian noise to graph-encoded user and item embeddings and a reverse denoising process generates two semantically informative views for contrastive learning. The framework also aligns visual and textual representations with the stable ID embedding, and augments item embeddings with an item-item graph built from top-K cosine neighbors. The authors report that on the Baby, Video, and Sports Amazon subsets, DiffCL achieves the best Recall@10/20 and NDCG@10/20 among all compared general and multimodal recommenders, beating the strongest multimodal baselines by roughly 3 to 8 percent. The contribution matters because it attacks two known weaknesses of multimodal recommendation—noise in self-supervised augmentation and cross-modal semantic inconsistency—without requiring new interaction data.

What carries the argument

The central mechanism is the diffusion graph contrastive learning module: a forward Markov chain $q(x_t|x_{t-1})=\mathcal{N}(x_t;\sqrt{1-\beta_t}x_{t-1},\beta_t I)$ destroys the graph-encoded visual and textual embeddings, while a learned reverse chain $p_\theta(x_{t-1}|x_t)$ produces denoised pseudo-features, and two such pseudo-features are used as positive pairs for InfoNCE. The claim is that these diffusion-built views preserve recommendation-relevant information better than random Gaussian or dropout perturbations. Two supporting mechanisms carry the rest: an ID-guided cross-modal alignment that minimizes Gaussian mean and variance differences between visual, textual, and ID embeddings, and an item-item graph built from normalized top-$K$ cosine similarities that aggregates neighbor item embeddings to relieve data sparsity.

What would settle it

Replace the diffusion module with a matched one-step Gaussian perturbation of the same variance (the forward schedule evaluated at a single early timestep, with no reverse network), keeping all other losses and hyperparameters fixed; if Recall@20 and NDCG@20 do not drop, the diffusion-specific denoising mechanism is not what produces the reported gains.

Watch

Extended reading notes

Core claim

On its own terms, DiffCL claims that multimodal recommendation can be improved by replacing the usual view-construction step in graph contrastive learning with a diffusion process. Given a graph-encoded embedding $E_m$ for modality $m \in \{v,t\}$, the forward chain $q(x_t | x_{t-1}) = \mathcal{N}(x_t; \sqrt{1-\beta_t}x_{t-1}, \beta_t I)$ corrupts the feature, and a parameterized reverse chain $p_\theta(x_{t-1}|x_t)$ denoises it; running this twice yields two views $E^1_m$ and $E^2_m$ that are pulled together with an InfoNCE loss. A second task aligns visual and textual feature distributions to the ID embedding by matching Gaussian means and variances, $L_{\text{align}} = |\mu_{\text{id}}-\mu_v| + |\sigma_{\text{id}}-\sigma_v| + |\mu_{\text{id}}-\mu_t| + |\sigma_{\text{id}}-\sigma_t|$. A third component builds a normalized item-item graph from top-$K$ cosine-similarity neighbors and aggregates neighbor item embeddings to strengthen the item representation. The paper reports that the combined loss reaches the best Recall@10/20 and NDCG@10/20 on Baby, Video, and Sports, and that ablations show each component contributes to the result.

Load-bearing premise

The load-bearing premise is that a forward-reverse diffusion process with learnable denoising networks, trained only by the downstream contrastive and ranking losses rather than by a separate reconstruction objective, produces two contrastive views that are both distinct and informative enough to improve recommendations; the paper does not specify the denoising network's training loss, timestep count, or noise schedule parameters.

Editorial extensions

If this is right

  • According to the paper, adding diffusion-built contrastive views alone improves Recall@20 over the no-diffusion baseline on all three datasets.
  • The paper reports that every two-component combination outperforms every single-component version, and the full DiffCL is best on all datasets.
  • The ID-guided alignment loss is claimed to enforce cross-modal semantic consistency without disrupting historical interaction information, since it matches Gaussian moments against the stable ID embedding rather than applying direct inter-modal losses.
  • The item-item graph component is reported to mitigate data sparsity; each of the three datasets is more than 99.8 percent sparse.
  • If the framework is correct, the largest gains should appear where multimodal content carries strong preference signal, as in the Video subset where improvements over the next best model range from 3.60 percent to 8.64 percent.

Reading between the lines

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

  • Beyond the paper's comparisons, the diffusion module could be tested against a one-step Gaussian perturbation with a matched noise budget; if the gains vanish, the denoising dynamics specifically, not the added noise level, drive the result.
  • An implication the authors leave implicit is that the same diffusion plus alignment recipe could transfer to cold-start and lower-density settings, where sparsity is even more severe than the 99.9-percent sparsity reported here.
  • A further testable extension is to ablate the ID embedding used for alignment with a randomly re-initialized embedding; if the gain persists, the stability and uniqueness of the ID embedding are not the active mechanism.
  • The reliance on fixed pre-extracted visual and textual features suggests the observed gains may be bounded by feature-encoder quality, so joint end-to-end training of encoders with the diffusion and alignment losses is a natural next experiment.
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 / 6 minor

Summary. The paper proposes DiffCL, a multimodal recommendation framework that combines graph convolutional encoders with a diffusion-based graph contrastive learning module, an item-item graph for feature enhancement, and an ID-guided cross-modal semantic alignment loss. The final objective is a weighted combination of a contrastive loss, an alignment loss, a BPR ranking loss, and L2 regularization. Experiments on three Amazon subsets (Baby, Video, Sports) report consistent improvements in Recall@10/20 and NDCG@10/20 over several general and multimodal recommender baselines, with ablations and hyperparameter sensitivity studies.

Significance. The core idea of using a diffusion process to generate contrastive views rather than random noise or edge dropout is timely and potentially valuable for multimodal recommendation. The empirical protocol is a strength: the authors use three standard datasets, standardized MMRec implementations of baselines, comparisons with recent diffusion-based and graph-based methods, and component ablations. If the missing training objectives and implementation details are supplied, and the reported ablation numbers are reconciled with the text, the framework could be a useful contribution. As submitted, however, the central diffusion module is underspecified to the point that the reported gains cannot currently be attributed to diffusion-based contrastive learning.

major comments (4)
  1. [IV-C, Eqs. (6)-(12) and Eq. (32)] The reverse process pθ(xt−1|xt) is defined with learnable networks for μθ and Σθ, but the total loss in Eq. (32) contains no denoising objective: there is no x0-prediction term, no ε-prediction term, and no variational bound that trains these networks to reverse the forward noising process. The only training signal for the module is the contrastive loss Lcl on the produced views E1 and E2. As written, the module is therefore not a diffusion model in the DDPM sense, and the statement in Section IV-C that it 'restore[s] the original interaction by predicting the original state' is unsupported. Please specify the denoising loss, the architecture of the μθ/Σθ networks, the number of steps T, the noise-schedule parameters (γmin, γmax, s), and how E1 and E2 are sampled from the forward and reverse chains; without these details the diffusion ablation in Table III cannot be interpreted.
  2. [IV-D3, Eqs. (25)-(27)] The 'Gaussian parameterization' of Eid, Ev, and Et is stated only as notation; the manuscript never defines how μid, σid, μv, σv, μt, and σt are computed from the corresponding embedding matrices. These quantities could be computed across feature dimensions, across node sets, or via learned projection heads, and the choice changes the alignment loss. Since Lalign is part of the total objective and the ablation DiffCLalign is used to claim the benefit of the alignment component, this omission makes the method non-implementable and the ablation unverifiable. Please provide the exact computation, state the dimensionality assumptions, and justify why learned ID embeddings, which are updated during training, are treated as stable.
  3. [Table III and Section V-C] The text claims that each component improves the system and that combinations of components yield further gains, but Table III contradicts this in two places. On Baby, the full DiffCL (R@20=0.0987, N@20=0.0433) is not better than DiffCLalign+h (R@20=0.0993, N@20=0.0432). On Sports, the full model (R@20=0.1095) is worse than DiffCLalign+h (R@20=0.1114), even though its N@20 is slightly higher. The ablation therefore does not support the conclusion that the full model is uniformly superior or that every component always contributes. Please correct the reported numbers, provide per-seed statistics, or revise the conclusions to match the table.
  4. [V, Tables II and III] All experimental results are reported as point estimates without error bars, confidence intervals, or significance tests, and the number of random seeds is not stated. The central SOTA claim rests partly on improvements of 3-5% over the best baseline in Table II, which is within the typical run-to-run variation for recommender models when seed effects are unreported. Please report means and standard deviations over at least three seeds and, if possible, paired significance tests for the ablation comparisons.
minor comments (6)
  1. [Eqs. (8)-(10)] The same symbol γt is used for 1−βt and for the cumulative product in Eq. (9); this should be written as γt and γ̄t, and Eq. (10) should state explicitly whether the multiplier of x0 is γt or γ̄t.
  2. [Eq. (3)] The graph convolution update in Eq. (3) does not specify how the aggregation is applied separately to users and items, and the notation on the right-hand side should index the node whose embedding is being updated. Please rewrite the equation for E_u,m and E_i,m explicitly.
  3. [V-A3] The three loss-weight tuples {0.1, 0.4, 0.7}, {0.01, 1.0, 1.0}, and {0.7, 0.4, 0.9} are not explicitly mapped to (λcl, λalign, λE); please state the order.
  4. [Figures 3-5 and Eq. (17)] Figure 3 uses λdiff on the x-axis while Eq. (17) defines λcl, and the x-axis label in Figure 4 appears inconsistent with the text; please harmonize the notation throughout.
  5. [V-C] The ablation descriptions refer to the alignment task as 'intra-modal semantic alignment,' but the method aligns textual and visual features through ID guidance; this should be 'cross-modal' or 'inter-modal' alignment.
  6. [Eq. (22)] The aggregation formula A(l)m uses A(l−1)jm without a comma separating indices; it should be written as A(l−1)j,m, and the normalized adjacency matrix should be denoted consistently as \S^m_{i,j}.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DiffCL's claims are benchmark-validated and its components are not derived from their own outputs.

full rationale

The recommendation performance claims are supported by held-out test evaluation against external baselines on three Amazon subsets, so the central results are not equivalent to the training objective. The two self-citations in the introduction ([6], [7]) are background examples of deep learning progress and are not load-bearing for any derivation. The diffusion graph contrastive module is trained end-to-end with the same contrastive loss that defines its views; while this means the views are not produced by an independently trained generative model, it is a standard self-supervised setup and does not make the reported test metrics forced by construction. The most substantial concern is under-specification, not circularity: Eqs. (6)-(12) define a forward/reverse diffusion process, but the total loss in Eq. (32) contains no explicit denoising or noise-prediction term, so the reverse-process networks are not trained to 'restore the original interaction' as claimed in Sec. IV-C. That is a missing training objective that impedes reproduction and verification, but it does not reduce any prediction to its input; it is a completeness issue rather than a circular-derivation issue. No other pattern (fitted input called prediction, imported uniqueness theorem, ansatz smuggled via citation, or renaming) is present. Score 0 reflects the absence of circularity; the under-specification should be handled as a correctness/reproducibility risk.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The model relies on several domain assumptions and ad hoc design choices. The most consequential free parameters are the per-dataset loss weights and the unreported diffusion schedule and GCN depth. No new entities are postulated.

free parameters (8)
  • loss weights (λ_cl, λ_align, λ_E) = Baby {0.1,0.4,0.7}, Video {0.01,1.0,1.0}, Sports {0.7,0.4,0.9} (order not explicitly stated)
    Section V-A3 reports three per-dataset tuples for the three loss terms; the tuple-to-term mapping is ambiguous.
  • InfoNCE temperature τ = 0.4
    Section V-A3.
  • KNN neighbor count K = 10
    Section IV-D1 states 'We fix K = 10.'
  • Initial fusion weight μ = 0.5 (trainable)
    Section IV-D2.
  • Learning rate = 0.001
    Section V-A3.
  • Dropout rate = 0.5
    Section V-A3.
  • Diffusion steps T and noise schedule parameters (γ_min, γ_max, s) = Not reported
    Equation (11) defines a linear scheduler but no values are given for any dataset.
  • Number of GCN layers L = Not reported
    Equation (4) sums over L layers, but L is never stated.
assumptions (5)
  • domain assumption Pre-extracted ResNet50 and BERT features carry meaningful semantics for recommendation.
    Used as the frozen input features E_v and E_t (Section IV-B); no fine-tuning.
  • domain assumption KNN graphs built on raw features with K=10 capture useful item-item relations.
    Section IV-D1; no sensitivity analysis for K beyond the fixed value.
  • ad hoc to paper Gaussian parameterization of embeddings and matching their mean/std to the ID distribution achieves cross-modal semantic alignment.
    Section IV-D3 asserts alignment without evidence that first and second moments are sufficient.
  • ad hoc to paper A forward-reverse diffusion process on embeddings, trained without an explicit denoising loss, produces informative contrastive views.
    Section IV-C; no training objective or architecture for p_theta is given.
  • domain assumption Contrastive InfoNCE loss on two diffusion-generated views improves representation robustness.
    Standard SSL assumption (Section IV-C3); not validated separately from end-task metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DiffCL: A Diffusion-Based Contrastive Learning Framework with Semantic Alignment for Multimodal Recommendations." pith.science (2026). https://pith.science/paper/M4XGPNJQ

@misc{pith2026250101066,
  author       = {Pith},
  title        = {Pith review of: DiffCL: A Diffusion-Based Contrastive Learning Framework with Semantic Alignment for Multimodal Recommendations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M4XGPNJQ}},
  note         = {Machine review of arXiv:2501.01066}
}
read the original abstract

Multimodal recommendation systems integrate diverse multimodal information into the feature representations of both items and users, thereby enabling a more comprehensive modeling of user preferences. However, existing methods are hindered by data sparsity and the inherent noise within multimodal data, which impedes the accurate capture of users' interest preferences. Additionally, discrepancies in the semantic representations of items across different modalities can adversely impact the prediction accuracy of recommendation models. To address these challenges, we introduce a novel diffusion-based contrastive learning framework (DiffCL) for multimodal recommendation. DiffCL employs a diffusion model to generate contrastive views that effectively mitigate the impact of noise during the contrastive learning phase. Furthermore, it improves semantic consistency across modalities by aligning distinct visual and textual semantic information through stable ID embeddings. Finally, the introduction of the Item-Item Graph enhances multimodal feature representations, thereby alleviating the adverse effects of data sparsity on the overall system performance. We conduct extensive experiments on three public datasets, and the results demonstrate the superiority and effectiveness of the DiffCL.

Figures

Figures reproduced from arXiv: 2501.01066 by the authors.

Figure 1
Figure 1. Two methods for constructing graph contrastive learning: edge dropout [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overview architecture of our DiffCL framework. The DiffCL consists of three modules. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 4
Figure 4. The performance of the DiffCL under various [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: The performance of the DiffCL under various [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 28 canonical work pages

  1. [1]

    Came: Content- and context-aware music embedding for recommendation,

    D. Wang, X. Zhang, D. Yu, G. Xu, and S. Deng, “Came: Content- and context-aware music embedding for recommendation,” IEEE Transac- tions on Neural Networks and Learning Systems , vol. 32, no. 3, pp. 1375–1388, 2021

    work page 2021

  2. [2]

    Cipl: Counterfactual interactive policy learning to eliminate popularity bias for online rec- ommendation,

    Y . Zheng, J. Qin, P. Wei, Z. Chen, and L. Lin, “Cipl: Counterfactual interactive policy learning to eliminate popularity bias for online rec- ommendation,” IEEE Transactions on Neural Networks and Learning Systems, pp. 1–14, 2023

    work page 2023

  3. [3]

    Sparse enhanced network: An adversarial generation method for robust augmentation in sequential recommendation,

    J. Chen, G. Zou, P. Zhou, W. Yirui, Z. Chen, H. Su, H. Wang, and Z. Gong, “Sparse enhanced network: An adversarial generation method for robust augmentation in sequential recommendation,” Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 8, pp. 8283– 8291, 2024

    work page 2024

  4. [4]

    Neural col- laborative filtering,

    X. He, L. Liao, H. Zhang, L. Nie, X. Hu, and T.-S. Chua, “Neural col- laborative filtering,” in Proceedings of the 26th international conference on world wide web , 2017, pp. 173–182

    2017

  5. [5]

    Factorization machines,

    S. Rendle, “Factorization machines,” in 2010 IEEE International con- ference on data mining . IEEE, 2010, pp. 995–1000

    work page 2010

  6. [6]

    Multi-scale conformer fusion network for multi-participant behavior analysis,

    Q. Song, R. Dian, B. Sun, J. Xie, and S. Li, “Multi-scale conformer fusion network for multi-participant behavior analysis,” in Proceedings of the 31st ACM International Conference on Multimedia , 2023, pp. 9472–9476

    work page 2023

  7. [7]

    Multimodal sparse transformer network for audio-visual speech recognition,

    Q. Song, B. Sun, and S. Li, “Multimodal sparse transformer network for audio-visual speech recognition,” IEEE Transactions on Neural Networks and Learning Systems , vol. 34, no. 12, pp. 10 028–10 038, 2023

    work page 2023

  8. [8]

    Adaptive deep modeling of users and items using side information for recommendation,

    J. Han, L. Zheng, Y . Xu, B. Zhang, F. Zhuang, P. S. Yu, and W. Zuo, “Adaptive deep modeling of users and items using side information for recommendation,” IEEE Transactions on Neural Networks and Learning Systems, vol. 31, no. 3, pp. 737–748, 2020

    work page 2020

Show all 51 references
  1. [9]

    Neural graph collaborative filtering,

    X. Wang, X. He, M. Wang, F. Feng, and T.-S. Chua, “Neural graph collaborative filtering,” in Proceedings of the 42nd international ACM SIGIR conference on Research and development in Information Re- trieval, 2019, pp. 165–174

    2019

  2. [10]

    Social attentive deep q- networks for recommender systems,

    Y . Lei, Z. Wang, W. Li, H. Pei, and Q. Dai, “Social attentive deep q- networks for recommender systems,” IEEE Transactions on Knowledge and Data Engineering , vol. 34, no. 5, pp. 2443–2457, 2020

    2020

  3. [11]

    Self-training enhanced: Network embedding and overlapping community detection with adversarial learning,

    J. Chen, Z. Gong, J. Mo, W. Wang, W. Wang, C. Wang, X. Dong, W. Liu, and K. Wu, “Self-training enhanced: Network embedding and overlapping community detection with adversarial learning,” IEEE Transactions on Neural Networks and Learning Systems, vol. 33, no. 11, pp. 6737–6748, 2022

    2022

  4. [12]

    Crl: Collaborative representation learning by coordinating topic modeling and network embeddings,

    J. Chen, Z. Gong, W. Wang, W. Liu, and X. Dong, “Crl: Collaborative representation learning by coordinating topic modeling and network embeddings,” IEEE Transactions on Neural Networks and Learning Systems, vol. 33, no. 8, pp. 3765–3777, 2022

    2022

  5. [13]

    Tag: Joint triple-hierarchical attention and gcn for review-based social recommender system,

    P. Qiao, Z. Zhang, Z. Li, Y . Zhang, K. Bian, Y . Li, and G. Wang, “Tag: Joint triple-hierarchical attention and gcn for review-based social recommender system,” IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 10, pp. 9904–9919, 2022

    2022

  6. [14]

    Enhancing social recommendation with adversarial graph convolutional networks,

    J. Yu, H. Yin, J. Li, M. Gao, Z. Huang, and L. Cui, “Enhancing social recommendation with adversarial graph convolutional networks,” IEEE Transactions on knowledge and data engineering , vol. 34, no. 8, pp. 3727–3739, 2020

    2020

  7. [15]

    Xsimgcl: Towards extremely simple graph contrastive learning for recommenda- tion,

    J. Yu, X. Xia, T. Chen, L. Cui, N. Q. V . Hung, and H. Yin, “Xsimgcl: Towards extremely simple graph contrastive learning for recommenda- tion,” IEEE Transactions on Knowledge and Data Engineering , vol. 36, no. 2, pp. 913–926, 2023

    2023

  8. [16]

    Mmgcn: Multi-modal graph convolution network for personalized recommenda- tion of micro-video,

    Y . Wei, X. Wang, L. Nie, X. He, R. Hong, and T.-S. Chua, “Mmgcn: Multi-modal graph convolution network for personalized recommenda- tion of micro-video,” in Proceedings of the 27th ACM international conference on multimedia , 2019, pp. 1437–1445

    2019

  9. [17]

    Multi-view graph convolutional network for multimedia recommendation,

    P. Yu, Z. Tan, G. Lu, and B.-K. Bao, “Multi-view graph convolutional network for multimedia recommendation,” in Proceedings of the 31st ACM International Conference on Multimedia , 2023, pp. 6576–6585. 11

    2023

  10. [18]

    Dualgnn: Dual graph neural network for multimedia recommendation,

    Q. Wang, Y . Wei, J. Yin, J. Wu, X. Song, and L. Nie, “Dualgnn: Dual graph neural network for multimedia recommendation,” IEEE Transactions on Multimedia , vol. 25, pp. 1074–1084, 2021

    2021

  11. [19]

    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,” IEEE Transactions on Knowledge and Data Engineering , vol. 36, no. 1, pp. 335–355, 2023

    2023

  12. [20]

    Improving graph collaborative filtering with neighborhood-enriched contrastive learning,

    Z. Lin, C. Tian, Y . Hou, and W. X. Zhao, “Improving graph collaborative filtering with neighborhood-enriched contrastive learning,” in Proceed- ings of the ACM web conference , 2022, pp. 2320–2329

    2022

  13. [21]

    Hyper- graph contrastive collaborative filtering,

    L. Xia, C. Huang, Y . Xu, J. Zhao, D. Yin, and J. Huang, “Hyper- graph contrastive collaborative filtering,” in Proceedings of the 45th International ACM SIGIR conference on research and development in information retrieval, 2022, pp. 70–79

    2022

  14. [22]

    Self- supervised graph learning for recommendation,

    J. Wu, X. Wang, F. Feng, X. He, L. Chen, J. Lian, and X. Xie, “Self- supervised graph learning for recommendation,” in Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval , 2021, pp. 726–735

    2021

  15. [23]

    Self- supervised multi-channel hypergraph convolutional network for social recommendation,

    J. Yu, H. Yin, J. Li, Q. Wang, N. Q. V . Hung, and X. Zhang, “Self- supervised multi-channel hypergraph convolutional network for social recommendation,” in Proceedings of the web conference 2021 , 2021, pp. 413–424

    2021

  16. [24]

    Bootstrap latent representations for multi-modal recommen- dation,

    X. Zhou, H. Zhou, Y . Liu, Z. Zeng, C. Miao, P. Wang, Y . You, and F. Jiang, “Bootstrap latent representations for multi-modal recommen- dation,” in Proceedings of the ACM Web Conference 2023 , 2023, pp. 845–854

    2023

  17. [25]

    Multi-modal graph contrastive learning for micro-video recommendation,

    Z. Yi, X. Wang, I. Ounis, and C. Macdonald, “Multi-modal graph contrastive learning for micro-video recommendation,” in Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2022, pp. 1807–1811

    2022

  18. [26]

    Multi-modal self-supervised learning for recommendation,

    W. Wei, C. Huang, L. Xia, and C. Zhang, “Multi-modal self-supervised learning for recommendation,” in Proceedings of the ACM Web Confer- ence, 2023, pp. 790–800

    2023

  19. [27]

    Learning image and user features for recommendation in social networks,

    X. Geng, H. Zhang, J. Bian, and T.-S. Chua, “Learning image and user features for recommendation in social networks,” in Proceedings of the IEEE international conference on computer vision , 2015, pp. 4274– 4282

    2015

  20. [28]

    Atten- tive collaborative filtering: Multimedia recommendation with item-and component-level attention,

    J. Chen, H. Zhang, X. He, L. Nie, W. Liu, and T.-S. Chua, “Atten- tive collaborative filtering: Multimedia recommendation with item-and component-level attention,” in Proceedings of the 40th International ACM SIGIR conference on Research and Development in Information Retrieva...

    2017

  21. [29]

    Collaborative knowledge base embedding for recommender systems,

    F. Zhang, N. J. Yuan, D. Lian, X. Xie, and W.-Y . Ma, “Collaborative knowledge base embedding for recommender systems,” in Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining , 2016, pp. 353–362

    2016

  22. [30]

    Mining latent structures for multimedia recommendation,

    J. Zhang, Y . Zhu, Q. Liu, S. Wu, S. Wang, and L. Wang, “Mining latent structures for multimedia recommendation,” in Proceedings of the 29th ACM international conference on multimedia , 2021, pp. 3872–3880

    2021

  23. [31]

    Diversifying collaborative filtering via graph spreading network and selective sam- pling,

    Y . Fang, H. Wu, Y . Zhao, L. Zhang, S. Qin, and X. Wang, “Diversifying collaborative filtering via graph spreading network and selective sam- pling,” IEEE Transactions on Neural Networks and Learning Systems , vol. 35, no. 10, pp. 13 860–13 873, 2024

    2024

  24. [32]

    Gcrec: Graph-augmented capsule network for next-item recommendation,

    B. Wu, X. He, Q. Zhang, M. Wang, and Y . Ye, “Gcrec: Graph-augmented capsule network for next-item recommendation,” IEEE Transactions on Neural Networks and Learning Systems , vol. 34, no. 12, pp. 10 164– 10 177, 2023

    2023

  25. [33]

    Joint multi-grained popularity-aware graph convolution collaborative filtering for recom- mendation,

    K. Liu, F. Xue, X. He, D. Guo, and R. Hong, “Joint multi-grained popularity-aware graph convolution collaborative filtering for recom- mendation,” IEEE Transactions on Computational Social Systems , vol. 10, no. 1, pp. 72–83, 2022

    2022

  26. [34]

    Diffusion models in vision: A survey,

    F.-A. Croitoru, V . Hondru, R. T. Ionescu, and M. Shah, “Diffusion models in vision: A survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 9, pp. 10 850–10 869, 2023

    2023

  27. [35]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems , vol. 33, pp. 6840– 6851, 2020

    2020

  28. [36]

    Bilateral denoising diffusion models,

    M. W. Lam, J. Wang, R. Huang, D. Su, and D. Yu, “Bilateral denoising diffusion models,” arXiv preprint arXiv:2108.11514 , 2021

    2021 arXiv

  29. [37]

    Plug- in diffusion model for sequential recommendation,

    H. Ma, R. Xie, L. Meng, X. Chen, X. Zhang, L. Lin, and Z. Kang, “Plug- in diffusion model for sequential recommendation,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 8, 2024, pp. 8886–8894

    2024

  30. [38]

    Generate what you prefer: Reshaping sequential recommendation via guided diffusion,

    Z. Yang, J. Wu, Z. Wang, X. Wang, Y . Yuan, and X. He, “Generate what you prefer: Reshaping sequential recommendation via guided diffusion,” Advances in Neural Information Processing Systems , vol. 36, 2024

    2024

  31. [39]

    Diffusion recommender model,

    W. Wang, Y . Xu, F. Feng, X. Lin, X. He, and T.-S. Chua, “Diffusion recommender model,” in Proceedings of the 46th International ACM SI- GIR Conference on Research and Development in Information Retrieval, 2023, pp. 832–841

    2023

  32. [40]

    Ld4mrec: Simplifying and pow- ering diffusion model for multimedia recommendation,

    P. Yu, Z. Tan, G. Lu, and B.-K. Bao, “Ld4mrec: Simplifying and pow- ering diffusion model for multimedia recommendation,” arXiv preprint arXiv:2309.15363, 2023

    2023 arXiv

  33. [41]

    Diffmm: Multi-modal diffusion model for recommendation,

    Y . Jiang, L. Xia, W. Wei, D. Luo, K. Lin, and C. Huang, “Diffmm: Multi-modal diffusion model for recommendation,” arXiv preprint arXiv:2406.11781, 2024

    2024 arXiv

  34. [42]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” arXiv preprint arXiv:1609.02907 , 2016

    2016 arXiv

  35. [43]

    Multimodal machine learning: A survey and taxonomy,

    T. Baltru ˇsaitis, C. Ahuja, and L.-P. Morency, “Multimodal machine learning: A survey and taxonomy,” IEEE transactions on pattern anal- ysis and machine intelligence , vol. 41, no. 2, pp. 423–443, 2018

    2018

  36. [44]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

    2016

  37. [45]

    Momentum contrast for unsupervised visual representation learning,

    K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Momentum contrast for unsupervised visual representation learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 9729–9738

    2020

  38. [46]

    Fed- erated probabilistic preference distribution modelling with compactness co-clustering for privacy-preserving multi-domain recommendation

    W. Liu, C. Chen, X. Liao, M. Hu, J. Yin, Y . Tan, and L. Zheng, “Fed- erated probabilistic preference distribution modelling with compactness co-clustering for privacy-preserving multi-domain recommendation.” in IJCAI, 2023, pp. 2206–2214

    2023

  39. [47]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, “Bert: Pre-training of deep bidirectional transformers for language understanding,” arXiv preprint arXiv:1810.04805 , 2018

    2018 arXiv

  40. [48]

    Bpr: Bayesian personalized ranking from implicit feedback,

    S. Rendle, C. Freudenthaler, Z. Gantner, and L. Schmidt-Thieme, “Bpr: Bayesian personalized ranking from implicit feedback,” arXiv preprint arXiv:1205.2618, 2012

    2012 arXiv

  41. [49]

    Self-supervised learning for multimedia recommendation,

    Z. Tao, X. Liu, Y . Xia, X. Wang, L. Yang, X. Huang, and T.-S. Chua, “Self-supervised learning for multimedia recommendation,” IEEE Transactions on Multimedia , 2022

    2022

  42. [50]

    A tale of two graphs: Freezing and denoising graph structures for multimodal recommendation,

    X. Zhou and Z. Shen, “A tale of two graphs: Freezing and denoising graph structures for multimodal recommendation,” in Proceedings of the 31st ACM International Conference on Multimedia , 2023, pp. 935–943

    2023

  43. [51]

    A comprehensive survey on multimodal recommender systems: Taxonomy, evaluation, and future directions,

    H. Zhou, X. Zhou, Z. Zeng, L. Zhang, and Z. Shen, “A comprehensive survey on multimodal recommender systems: Taxonomy, evaluation, and future directions,” arXiv preprint arXiv:2302.04473 , 2023

    2023 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.