Pith. sign in

REVIEW 3 major objections 4 minor 38 references

Modality Alignment with Multi-scale Bilateral Attention for Multimodal Recommendation

T0 review · 3 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper proposes MambaRec, a multimodal recommender that aligns visual and textual item features at two levels—local detail through a dilated attention module, global distribution through MMD and contrastive losses—and reports top Recall

desk verdict Plausible engineering combination with consistent gains, but the core DREAM module is underspecified for 1D inputs to the point of being vacuous as described; needs major clarifications before the results can be trusted. read the letter →

arxiv 2509.09114 v1 pith:KO2R43AM submitted 2025-09-11 cs.IR

Kelin Ren , Chan-Yang Ju , Dong-Ho Lee This is my paper
classification cs.IR
keywords multimodalrecommendationmodalityalignmentdilatedconvolutionattentionmechanismMaximumMeanDiscrepancycontrastivelearningdimensionalityreductionrecommendersystems
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 tries to establish that multimodal recommendation systems get better when the visual and textual features of items are explicitly aligned at two complementary levels: fine-grained local patterns and global distribution-level consistency. Prior methods, the paper argues, fuse modalities without such alignment and therefore suffer from modal noise and representational bias. To test that claim, the paper builds MambaRec, whose local alignment is carried by the Dilated Refinement Attention Module (DREAM)—multi-scale dilated convolutions with channel and spatial attention—and whose global alignment is carried by Maximum Mean Discrepancy (MMD) plus InfoNCE contrastive losses, with a dimensionality reduction scheme to keep the cost down. On Baby, Sports, and Clothing subsets of Amazon reviews, MambaRec outperforms nine general and multimodal baselines on Recall@10/20 and NDCG@10/20 with reported significance at p<0.01, and ablations indicate both alignment levels contribute, with local alignment mattering more. If true, the claim is that two-level alignment, not fusion alone, is what drives gains in multimodal recommendation.

What carries the argument

Three mechanisms carry the argument. (1) The Dilated Refinement Attention Module (DREAM), a five-branch multi-scale block—1×1 convolution, dilated 3×3 convolutions with rates 6/12/18, and global average pooling upsampled back—whose concatenated output is weighted by a channel-attention branch and a spatial-attention branch, merged by element-wise maximum; it supplies the fine-grained local alignment that prior static fusion lacks. (2) A global alignment loss pairing Maximum Mean Discrepancy (MMD) with a Gaussian kernel against an InfoNCE contrastive term, pulling visual and textual feature distributions together; this is the paper's consistency mechanism. (3) A linear dimensionality reductio

What would settle it

Run MambaRec on the Baby or Sports dataset with DREAM replaced by a matched-parameter MLP of the same capacity (or with dilated convolutions removed but attention retained), keeping all other losses and hyperparameters identical. If Recall@20 and NDCG@20 do not drop materially, the local-alignment module is not the source of the reported gains. A complementary check removes the MMD loss while keeping InfoNCE: if performance is unchanged, the global distribution constraint adds nothing beyond contrastive learning. Both are single-variable changes a reader could implement directly.

Watch

Extended reading notes

Core claim

The paper claims prior multimodal recommenders fail on two fronts—weak fine-grained cross-modal association and missing global distribution-level consistency—and that fixing both jointly improves accuracy. Local alignment is carried by DREAM, a five-branch module (1×1 conv, three dilated 3×3 convs at rates 6/12/18, global pooling) fused and recalibrated by channel and spatial attention merged by element-wise max. Global alignment is carried by Gaussian-kernel Maximum Mean Discrepancy plus InfoNCE contrastive loss. On Baby, Sports, and Clothing, MambaRec reports best Recall and NDCG at K=10/20 over nine baselines (p<0.01); ablations show local alignment matters more than global, text beats vi

Load-bearing premise

The argument depends on the assumption that DREAM, which is defined for 2D image-style feature maps (channels × height × width), still performs meaningful 'local alignment' when applied to one-dimensional 64-dimensional embedding vectors; the paper never states how the vectors are reshaped or why dilated 2D convolutions and spatial attention are appropriate at that scale.

Editorial extensions

If this is right

  • If the results hold, two-level alignment—local attention-based matching plus distribution-level regularization—is a transferable recipe that other multimodal backbones could adopt without changing their user-item graph.
  • The reported single-modality results imply text features are the stronger modality on these datasets, so methods that weight or de-noise modalities asymmetrically could extract further gains.
  • The ablation ordering (local alignment more important than global) predicts that future improvements should target the local matching module rather than the distribution loss.
  • The reduction-factor finding (factor 8 peaks while more aggressive compression degrades) suggests a testable sweet spot between compressing away noise and losing discriminative detail.
  • If MMD truly reduces modal bias, MambaRec should generalize better to cold-start items with missing visual or textual features—something the paper motivates but does not directly test.

Reading between the lines

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

  • The name MambaRec evokes state-space sequence models, but the architecture described is convolutional attention plus loss regularization; nothing depends on sequential state-space machinery, so readers should judge it as a fusion-and-alignment method rather than a Mamba-style model.
  • DREAM is specified for 2D feature maps (C×H×W), yet the experiments apply it to 64-dimensional vectors after dimensionality reduction; the paper never explains the reshaping, so the module may act as a nonlinear feature transformer rather than literally aligning spatial regions. Isolating the dilated convolutions from the dual attention would show which ingredient carries the gain.
  • Because MMD and InfoNCE are modality-agnostic, the global-alignment half should transfer directly to other modality pairs (audio/video, image/audio) by swapping feature extractors—an untested extension the paper gestures at in its conclusion.
  • The t-SNE visualizations suggest alignment changes embedding geometry toward a more uniform spread, pointing to a testable side effect: better-calibrated score distributions and possibly better robustness under distribution shift, neither of which the paper measures directly.
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 / 4 minor

Summary. The paper proposes MambaRec, a multimodal recommendation framework that combines three components: (i) a Dilated Refinement Attention Module (DREAM) intended to align fine-grained visual and textual features via multi-scale dilated convolutions plus channel/spatial attention; (ii) global distribution alignment using Maximum Mean Discrepancy (MMD) and InfoNCE contrastive loss; and (iii) a linear dimensionality-reduction scheme to reduce memory overhead. The model is evaluated on three Amazon subsets (Baby, Sports, Clothing) against general and multimodal baselines, reporting improvements on Recall@10/20 and NDCG@10/20, with additional ablations, hyperparameter sensitivity, efficiency measurements, and t-SNE visualizations. The code is publicly available.

Significance. If the DREAM module operates as described and the empirical improvements are reproducible, the paper would offer a useful and practical combination of existing techniques (ASPP-style multi-scale convolution, CBAM-style attention, MMD, and InfoNCE) for multimodal recommendation, supported by experiments on three public datasets, ablations, and released code. However, the central architectural claim is currently undermined by a missing specification of how DREAM is applied to the actual 1D feature vectors, and the statistical significance claim in Table 2 is not backed by variance information. These issues must be resolved before the reported gains can be attributed to the proposed mechanism.

major comments (3)
  1. [§4.1, §4.3, Eqs. (12)–(13)] DREAM is formally defined for an input X∈R^{C×H×W}, but §4.3 states that the reduced features V′,T′∈R^{N×d} are 'further input into the DREAM module', and no reshape to a 2D spatial map is given. With d=48 (min(4096,384)/8 under Eq. (13)) or d=64, the input is a vector, not a feature map. If it is reshaped to [C,1,1], the 3×3 dilated convolutions with rates 6, 12, and 18 all collapse to 1×1 convolutions, and the spatial attention in Eqs. (4)–(5) operates on a single pixel. If reshaped to, e.g., 8×8, the rate-18 branch has a 37×37 effective receptive field, larger than the input, so most taps fall on zero padding. In either case the multi-scale bilateral attention described in Eqs. (1)–(6) is not actually implemented on the tensors the model consumes. This is load-bearing because the paper attributes the Table 2 gains and the Figure 3 ablation results to DREAM. Please specify the exact te
  2. [§5.2, Table 2] The caption of Table 2 states 'The t-tests verified the significance of performance improvements with p-value<0.01', but the table reports no standard deviations, no number of independent runs, and no actual p-values. §5.1.4 additionally states that the same random seed was used in all implementations, which would prevent estimating variance across seeds. Without this information, the claim of statistical significance is not verifiable, and several reported margins are small (e.g., Baby Recall@20: MambaRec 0.1013 vs. LGMRec 0.1002). Please report mean±std over multiple seeds and provide the test details (paired/one-sided, number of runs, p-values) or remove the significance claim.
  3. [§2, §5.1.2, Table 2] The related work highlights BM3 [35] as a recent self-supervised multimodal recommendation model, but BM3 is absent from the baseline list in §5.1.2 and from Table 2. Given the paper's claim of outperforming state-of-the-art multimodal methods on all three datasets, the omission of a directly relevant and widely used baseline is material. Please add BM3 (and, if feasible, other recently cited models such as MONET) to the comparison, or justify their exclusion explicitly.
minor comments (4)
  1. [§4] There are several typos and inconsistencies: 'MamabaRec' appears instead of 'MambaRec'; §4.2 heading reads 'Global Attribution Alignment' rather than 'Global Distribution Alignment'; 'comparative learning' should be 'contrastive learning'; and 'Alignement' appears in §4 opening. Please proofread.
  2. [§4.4] The 'Multi-View Encoder' is described only in prose with no equations, layer details, or input/output shapes. This limits reproducibility from the text; the public code helps, but the section should be expanded so that the architecture is self-contained.
  3. [§5.1.4, §5.4] The relation between the user/item embedding dimension (64) and the reduced modality dimension d (Eq. (13), which gives 48 when r=8) should be clarified. In addition, Figures 3, 5, and 6 lack numerical labels/error bars; please add axis units and quantitative annotations, and specify whether the y-axis in Figures 5–6 is Recall@20 or another metric.
  4. [§5.5, Figures 7–8] The visualization description is confusing: it says the Baby dataset is represented in red and Sports in blue, but the figures are labeled separately by dataset. Please clarify what is plotted in each panel (feature distributions of MambaRec vs. VBPR? combined datasets?) and how the kernel density estimate was computed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical architecture comparison with held-out evaluation; flagged DREAM reshape ambiguity is a correctness/validity issue, not a circular derivation.

full rationale

MambaRec is an empirical multimodal-recommendation paper whose claims are benchmark results, not first-principles derivations. Its components are a DREAM module assembled from externally cited building blocks (ASPP-style dilated convolutions [2], CBAM-style channel/spatial attention [28]), standard MMD [6] and InfoNCE [18] losses, a linear dimensionality reduction (Eq. 12), and BPR optimization (Eq. 15). No equation defines a predicted quantity in terms of a fitted constant or in terms of the target metric, so the self-definitional and fitted-input-called-prediction patterns do not apply. Hyperparameters are selected on a validation split and metrics are reported on a held-out test split (Section 5.1.3), so the evaluation is not forced by construction. The paper does not invoke any load-bearing self-citation: the prior work it builds on (DeepLabV3, CBAM, MMD, InfoNCE, MGCN, MMRec) is third-party or framework-level, and none of those citations is used to forbid alternatives or to uniquely justify the architecture. The reviewer-flagged ambiguity that DREAM is specified for 2D feature maps X∈R^{C×H×W} (Section 4.1) but receives reduced 1D vectors V',T'∈R^{N×d} (Section 4.3) is a real specification gap that could undermine the local-alignment claim; however, an underspecified or even vacuous module is a correctness/validity concern, not circularity, because the paper's empirical results do not logically reduce to its own assumptions. Accordingly, the appropriate circularity score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The central claim depends on several hand-chosen hyperparameters and the assumption that standard pretrained features and the evaluation protocol are adequate. No new physical entities are proposed; DREAM is an architectural component.

free parameters (6)
  • lambda_mmd (MMD loss weight) = 0.15 (tuned from [0.1, 0.15, 0.2])
    Selected by validation; the reported performance depends on this choice.
  • lambda_cl (InfoNCE loss weight) = 0.01 (from sensitivity analysis)
    Paper concludes this is optimal on Baby and Sports datasets.
  • Gaussian kernel bandwidth sigma = tuned from [1.0, 1.5, 2.0], final value not explicitly stated
    Kernel bandwidth affects MMD computation and is chosen by hand.
  • Reduction factor r = 8
    Chosen as the best trade-off between performance and resource use.
  • InfoNCE temperature tau = not reported
    Standard contrastive hyperparameter, not specified in the paper.
  • Dilation rates in DREAM = 6, 12, 18
    Hand-selected to cover medium and long range context.
assumptions (5)
  • standard math MMD with Gaussian kernel measures distributional difference in RKHS
    Standard kernel two-sample test theory; the paper relies on it without proof.
  • standard math InfoNCE approximates mutual information between paired modalities
    Common contrastive learning objective; taken as given.
  • domain assumption Pretrained VGG16 and Sentence-BERT features are informative for item similarity
    No analysis of feature quality; the model's input is trusted.
  • domain assumption Linear projection can unify heterogeneous modalities into a common space
    A single linear layer is used for each modality before alignment.
  • domain assumption Amazon 5-core subsets are a valid benchmark for multimodal recommendation
    The paper follows prior work in using these datasets; validity is assumed.
invented entities (1)
  • DREAM (Dilated Refinement Attention Module)
    purpose: Achieve local feature alignment via multi-scale dilated convolutions and dual attention
    A new architectural block, but its effectiveness is only demonstrated on the three datasets in this paper; no independent falsifiable handle outside these experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Modality Alignment with Multi-scale Bilateral Attention for Multimodal Recommendation." pith.science (2026). https://pith.science/paper/KO2R43AM

@misc{pith2026250909114,
  author       = {Pith},
  title        = {Pith review of: Modality Alignment with Multi-scale Bilateral Attention for Multimodal Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KO2R43AM}},
  note         = {Machine review of arXiv:2509.09114}
}
read the original abstract

Multimodal recommendation systems are increasingly becoming foundational technologies for e-commerce and content platforms, enabling personalized services by jointly modeling users' historical behaviors and the multimodal features of items (e.g., visual and textual). However, most existing methods rely on either static fusion strategies or graph-based local interaction modeling, facing two critical limitations: (1) insufficient ability to model fine-grained cross-modal associations, leading to suboptimal fusion quality; and (2) a lack of global distribution-level consistency, causing representational bias. To address these, we propose MambaRec, a novel framework that integrates local feature alignment and global distribution regularization via attention-guided learning. At its core, we introduce the Dilated Refinement Attention Module (DREAM), which uses multi-scale dilated convolutions with channel-wise and spatial attention to align fine-grained semantic patterns between visual and textual modalities. This module captures hierarchical relationships and context-aware associations, improving cross-modal semantic modeling. Additionally, we apply Maximum Mean Discrepancy (MMD) and contrastive loss functions to constrain global modality alignment, enhancing semantic consistency. This dual regularization reduces mode-specific deviations and boosts robustness. To improve scalability, MambaRec employs a dimensionality reduction strategy to lower the computational cost of high-dimensional multimodal features. Extensive experiments on real-world e-commerce datasets show that MambaRec outperforms existing methods in fusion quality, generalization, and efficiency. Our code has been made publicly available at https://github.com/rkl71/MambaRec.

Figures

Figures reproduced from arXiv: 2509.09114 by the authors.

Figure 1
Figure 1. (a) illustrates the process in which existing methods [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overall architecture of the proposed MambaRec, comprising three key components. (i) The local alignment [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Ablation studies on the proposed MambaRec [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Variation of MambaRec with 𝜆𝑐𝑙 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Variation of MambaRec with 𝜆𝑚𝑚𝑑 5.4.2 Effects of the weight of 𝜆𝑚𝑚𝑑 . We compared Recall@20 per￾formance on Baby and Sports datasets under different combina￾tions of reduction factors and maximum mean discrepancy weight (𝜆𝑚𝑚𝑑 ), as shown in [PITH_FULL_IMAGE:figures/fu…
Figure 6
Figure 6. Figure 6: Impact of Reduction Factor [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Distribution of representations for Baby Datasets [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Distribution of representations for Sports Datasets [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 10 linked inside Pith

  1. [35]

    Xin Zhou, Hongyu Zhou, Yong Liu, Zhiwei Zeng, Chunyan Miao, Pengwei Wang, Yuan You, and Feijun Jiang. 2023. Bootstrap latent representations for multi- modal recommendation. InProceedings of the ACM web conference 2023. 845–854

    2023

  2. [1]

    Jie Cai, Xin Wang, Haoyang Li, Ziwei Zhang, and Wenwu Zhu. 2024. Multimodal graph neural architecture search under distribution shifts. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 8227–8235

    2024

  3. [2]

    Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. 2017. Rethinking atrous convolution for semantic image segmentation.arXiv preprint arXiv:1706.05587(2017)

    arXiv 2017

  4. [3]

    Yu Cheng, Duo Wang, Pan Zhou, and Tao Zhang. 2017. A survey of model compression and acceleration for deep neural networks.arXiv preprint arXiv:1710.09282(2017)

    arXiv 2017

  5. [4]

    Jun Gao, Di He, Xu Tan, Tao Qin, Liwei Wang, and Tie-Yan Liu. 2019. Repre- sentation degeneration problem in training natural language generation models. arXiv preprint arXiv:1907.12009(2019)

    arXiv 2019

  6. [5]

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

    2010

  7. [6]

    Arthur Gretton, Karsten M Borgwardt, Malte J Rasch, Bernhard Schölkopf, and Alexander Smola. 2012. A kernel two-sample test.The Journal of Machine Learning Research13, 1 (2012), 723–773

    2012

  8. [7]

    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

Show all 38 references
  1. [8]

    Song Han, Huizi Mao, and William J Dally. 2015. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149(2015)

    2015 arXiv

  2. [9]

    Ruining He and Julian McAuley. 2016. VBPR: visual bayesian personalized ranking from implicit feedback. InProceedings of the AAAI conference on artificial intelligence, Vol. 30

    2016

  3. [10]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. InProceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval. 639–648

    2020

  4. [11]

    Yungi Kim, Taeri Kim, Won-Yong Shin, and Sang-Wook Kim. 2024. MONET: Modality-embracing graph convolutional network and target-aware attention for multimedia recommendation. InProceedings of the 17th ACM International Conference on Web Search and Data Mining. 332–340

    2024

  5. [12]

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

    2014 arXiv

  6. [13]

    Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. 2021. Align before fuse: Vision and language repre- sentation learning with momentum distillation.Advances in neural information processing systems34 (2021), 9694–9705

    2021

  7. [14]

    Qimai Li, Zhichao Han, and Xiao-Ming Wu. 2018. Deeper insights into graph convolutional networks for semi-supervised learning. InProceedings of the AAAI conference on artificial intelligence, Vol. 32

    2018

  8. [15]

    Defu Lian, Haoyu Wang, Zheng Liu, Jianxun Lian, Enhong Chen, and Xing Xie. 2020. Lightrec: A memory and search-efficient recommender system. In Proceedings of The Web Conference 2020. 695–705

    2020

  9. [16]

    Kang Liu, Feng Xue, Dan Guo, Peijie Sun, Shengsheng Qian, and Richang Hong

  10. [17]

    Qidong Liu, Jiaxi Hu, Yutian Xiao, Xiangyu Zhao, Jingtong Gao, Wanyu Wang, Qing Li, and Jiliang Tang. 2024. Multimodal Recommender Systems: A Survey. ACM Comput. Surv.57, 2, Article 26 (Oct. 2024), 17 pages. doi:10.1145/3695461

    2024 doi

  11. [18]

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

    2018 arXiv

  12. [19]

    Ruihong Qiu, Zi Huang, Hongzhi Yin, and Zijian Wang. 2022. Contrastive learning for representation degeneration problem in sequential recommendation. InProceedings of the fifteenth ACM international conference on web search and data mining. 813–823

    2022

  13. [20]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks.arXiv preprint arXiv:1908.10084(2019)

    2019 arXiv

  14. [21]

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

  15. [22]

    Karen Simonyan and Andrew Zisserman. 2014. Very deep convolutional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556(2014)

    2014 arXiv

  16. [23]

    Zhulin Tao, Xiaohao Liu, Yewei Xia, Xiang Wang, Lifang Yang, Xianglin Huang, and Tat-Seng Chua. 2022. Self-supervised learning for multimedia recommenda- tion.IEEE Transactions on Multimedia25 (2022), 5107–5116

    2022

  17. [24]

    George R Terrell and David W Scott. 1992. Variable kernel density estimation. The Annals of Statistics(1992), 1236–1265

    1992

  18. [25]

    Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-SNE. Journal of machine learning research9, 11 (2008)

    2008

  19. [26]

    Yinwei Wei, Xiang Wang, Liqiang Nie, Xiangnan He, and Tat-Seng Chua. 2020. Graph-refined convolutional network for multimedia recommendation with implicit feedback. InProceedings of the 28th ACM international conference on multimedia. 3541–3549

    2020

  20. [27]

    Yinwei Wei, Xiang Wang, Liqiang Nie, Xiangnan He, Richang Hong, and Tat-Seng Chua. 2019. MMGCN: Multi-modal graph convolution network for personalized recommendation of micro-video. InProceedings of the 27th ACM international conference on multimedia. 1437–1445

    2019

  21. [28]

    Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. 2018. Cbam: Convolutional block attention module. InProceedings of the European conference on computer vision (ECCV). 3–19

    2018

  22. [29]

    Shuo Yang, Zhaopan Xu, Kai Wang, Yang You, Hongxun Yao, Tongliang Liu, and Min Xu. 2023. Bicro: Noisy correspondence rectification for multi-modality data via bi-directional cross-modal similarity consistency. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat...

    2023

  23. [30]

    Penghang Yu, Zhiyi Tan, Guanming Lu, and Bing-Kun Bao. 2023. Multi-view graph convolutional network for multimedia recommendation. InProceedings of the 31st ACM international conference on multimedia. 6576–6585

    2023

  24. [31]

    Jinghao Zhang, Yanqiao Zhu, Qiang Liu, Shu Wu, Shuhui Wang, and Liang Wang

  25. [32]

    Hongyu Zhou, Xin Zhou, Zhiwei Zeng, Lingzi Zhang, and Zhiqi Shen. 2023. A comprehensive survey on multimodal recommender systems: Taxonomy, evalua- tion, and future directions.arXiv preprint arXiv:2302.04473(2023)

    2023 arXiv

  26. [33]

    Xin Zhou. 2023. Mmrec: Simplifying multimodal recommendation. InProceedings of the 5th ACM International Conference on Multimedia in Asia Workshops. 1–2

    2023

  27. [34]

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

    2023

  28. [2012]

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

    2012 arXiv

  29. [2021]

    InProceedings of the 29th ACM international conference on multimedia

    Mining latent structures for multimedia recommendation. InProceedings of the 29th ACM international conference on multimedia. 3872–3880

  30. [2023]

    Multimodal graph contrastive learning for multimedia-based recommenda- tion.IEEE Transactions on Multimedia25 (2023), 9343–9355

    2023

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.