MDE: Modality Discrimination Enhancement for Multi-modal Recommendation

Hang Zhou; Huijing Zhan; Yucheng Wang

REVIEW 4 major objections 5 minor 25 references

MDE: Modality Discrimination Enhancement for Multi-modal Recommendation

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read MDE claims that actively widening visual-textual differences, balanced node-by-node against alignment, improves multi-modal recommendation.

desk verdict Equation (6) is inverted: minimizing L_diff reduces modality differences instead of amplifying them, so the paper's central claim is unsupported as written. read the letter →

arxiv 2502.18481 v1 pith:N62WTK4R submitted 2025-02-08 cs.IR cs.AI

Hang Zhou , Yucheng Wang , Huijing Zhan This is my paper · ORCID

classification cs.IRcs.AI

keywords multi-modalrecommendationmodality-specificfeaturesmodality-sharedmodalitydifferenceamplificationnode-leveltrade-offcontrastivelearninggraphneuralnetworksBayesianpersonalizedranking

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

Multi-modal recommenders typically align the visual and textual representations of an item, which preserves what the modalities share but washes out what makes each modality informative on its own. This paper argues that recommendation accuracy improves when the model also does the opposite—actively amplifies differences between modality representations—and decides per user and per item how much differentiation versus alignment to apply. The proposed framework, MDE, pairs a modality difference amplification loss with a contrastive similarity alignment loss and couples them through a node-level trade-off driven by learned visual/textual preference weights. On the Baby and Sports benchmarks it reports gains over prior state-of-the-art models across recall, precision, MAP, and NDCG, and near-parity on Clothing. The paper's central claim is that jointly capturing modality-shared and modality-specific features, rather than alignment alone, is what drives the improvement.

What carries the argument

The load-bearing mechanism is the Node-Level Trade-off (NLT), a per-node pair of weight matrices built from learned modality preferences. For each item (and similarly for each user), the fusion module learns a scalar preference $P_{i,v}$ for the visual modality and $P_{i,t}$ for the textual modality; the trade-off then sets $W_{\mathrm{diff}} = f_b(|P_{i,v} - P_{i,t}|)$ and $W_{\mathrm{cl}} = f_b(1 - |P_{i,v} - P_{i,t}|)$, where $f_b$ broadcasts the scalar difference to the feature dimension. $W_{\mathrm{diff}}$ scales a loss that pushes visual and textual representations apart, while $W_{\mathrm{cl}}$ scales a contrastive loss that aligns them. Because one weight is the complement of the other, the mechanism automatically shifts each node's objective between difference amplification and alignment according to how strongly that node's modalities are predicted to disagree.

What would settle it

Measure the per-node preference gap $|P_{i,v} - P_{i,t}|$ on a validation set, then compare each node's recommendation-quality change when the difference-amplification loss is removed; if there is no positive correlation between gap size and gain from amplification, the Node-Level Trade-off hypothesis is false. A simpler check is to train MDE on a dataset where modality preferences are identical across users by construction and see whether it still beats a static-weight variant; if it does, the stated trade-off mechanism is not what drives the gains.

Watch

Extended reading notes

Core claim

The paper's central discovery is that explicit modality discrimination—encouraging visual and textual embeddings to differ—does not destroy the shared semantics needed for recommendation, provided the conflict is managed node by node. MDE learns a preference weight for each modality on each user and item, fuses features accordingly, and then uses the absolute preference gap to set complementary weights on a difference-amplification loss and an alignment loss. The result, on the paper's evidence, is that the model captures modality-specific cues such as shape and color versus function and price that pure alignment discards, yielding top-5 Recall@5 of 0.0414 on Baby and 0.0479 on Sports, up from 0.0374 and 0.0449 for DRAGON, the previous best in its tables, with matching gains on MAP and NDCG. On Clothing the recall is essentially tied (0.0402 vs 0.0401) while MAP and NDCG are slightly lower, which the paper attributes to unstable user modality preferences in that domain.

Load-bearing premise

The load-bearing premise is that a larger learned gap between a node's visual and textual preference weights really means differentiation is more useful for that node than alignment; if the true relationship is not monotone, the trade-off weights lose their meaning and the reported gains lose their stated cause.

Editorial extensions

If this is right

Because each user and item carries its own learned visual/textual preference, MDE treats modality importance as heterogeneous across nodes rather than as a single global weight.
Because difference amplification and alignment are combined in one objective, preserving modality-specific cues does not require abandoning shared semantic structure; the two losses coexist under the trade-off.
The ablation results imply that removing any of the three components—difference amplification, similarity alignment, or the node-level trade-off—reduces Recall@5, so each is load-bearing for the reported gain.
The Clothing results imply that MDE's advantage depends on the stability of user modality preferences; when those preferences are noisy, the trade-off helps less.

Reading between the lines

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

If the node-level trade-off hypothesis is right, the per-node gain from difference amplification should be concentrated on nodes with larger learned preference gaps; the paper does not report this, but it is a direct testable corollary.
On datasets where one modality dominates almost every user, the learned preference gaps become uniformly small or large, and the node-level trade-off should collapse toward a static weighting; this predicts where MDE's margin will shrink beyond the reported Clothing case.
The same amplify-differences-then-align recipe could transfer to other multi-modal tasks such as retrieval or classification that need modality-specific cues, but that transfer is not part of the paper's claims.

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.

Desk Editor's Note

read the letter

Quick take: the core mechanism is mathematically backwards. Equation (6) defines L_diff as a positive norm over the absolute difference between modality embeddings, so minimizing L_diff pushes those embeddings together. The text says the opposite. This is a load-bearing inconsistency, not a tuning issue, and it undercuts the entire MDA contribution.

What's worth keeping: the framing is honest and the problem is real — preserving modality-specific information in multimodal recommendation is a legitimate gap. The node-level trade-off (NLT) based on learned preference weights is a thoughtful way to balance alignment and differentiation, and the ablation design is clean. The paper also uses standard benchmarks and MMRec, which makes reproduction easier if code were included. On Clothing, the authors openly acknowledge the weaker result and try to explain it, which is more candid than most.

But the soft spots are serious beyond the sign error. There are no error bars or significance tests, and the reported gains are modest — around 10% relative on Baby, 7% on Sports, and effectively zero on Clothing, where MDE's MAP and NDCG are lower than DRAGON's. The NLT hypothesis is asserted without independent evidence; it may be a reasonable heuristic, but it's not validated. The paper also doesn't release code or hyperparameter details.

Given Eq. (6), the ablation results are difficult to interpret. If MDA as written aligns modalities, then the "w/o MDA" variant removes an alignment loss, and the comparison doesn't test the claimed difference amplification. This makes the internal evidence for the core idea even weaker.

Verdict: reject. The central claim is unsupported by the paper's own equations. With a corrected objective (e.g., a negative distance term or maximizing a distance), proper uncertainty reporting, and code, the NLT idea might be worth revisiting. As it stands, I wouldn't cite it or send it to review.

Referee Report

4 major / 5 minor

Summary. The paper proposes MDE, a multimodal recommendation framework that learns user and item representations from heterogeneous and homogeneous graphs, fuses visual and textual features using learnable node-level modality preferences, and introduces two auxiliary losses: Modality Difference Amplification (MDA) and Modality Similarity Alignment (MSA), balanced by a Node-Level Trade-off (NLT). The claimed contribution is that amplifying cross-modal differences while preserving shared semantics improves Top-5 recommendation accuracy. Experiments on Baby, Sports, and Clothing report gains over several baselines, with ablations for each module. The central technical problem is that the MDA loss in Eq. (6) is written so that its minimization reduces cross-modal differences, contradicting the paper's central claim.

Significance. If the mechanism worked as described, the node-level preference-based trade-off would be a simple and potentially useful addition to graph-based multimodal recommendation, and the paper's use of three Amazon benchmarks and MMRec baselines is appropriate. The paper also includes ablation and trade-off analyses, which are strengths. However, no code or reproducibility artifact is provided, and the mathematical formulation of the core MDA term is internally inconsistent with the stated purpose. Because the main empirical gains are attributed to that term, the results cannot be interpreted as evidence for the proposed mechanism in its current written form. The Clothing results also weaken the broad 'significantly outperforms' claim.

major comments (4)

[II-C, Eq. (6)] The loss L_diff is defined as σ_diff·||−|H*_{i,v}−H*_{i,t}|⊙W_diff||_2. Because the L2 norm of a vector equals the norm of its negative, minimizing L_diff minimizes the weighted absolute difference between the visual and textual embeddings. The sentence 'By minimizing L_diff, we can maximize the differences between modality features' is therefore mathematically false. Since Eq. (11) adds L_diff with a positive coefficient, the final objective as written can only shrink modality differences, not amplify them. This invalidates the central MDA mechanism and makes the 'w/o MDA' ablation in Table II uninterpretable as evidence for difference amplification.
[II-C, Eq. (7)] The final alignment loss is written as L_cl = σ_cl·(L_t→v_cl + L_t→v_cl), so the previously defined L_v→t_cl term never appears in the objective. This is an internal inconsistency: if the implementation uses both directional contrastive terms, the text should be corrected; if it does not, the MSA loss is not the one described. In either case, the optimization objective in the paper is ambiguous.
[Table I] The abstract and conclusion state that MDE 'significantly outperforms' state-of-the-art methods, but on Clothing MDE's Recall@5 is 0.0402 versus 0.0401 for DRAGON, while its MAP (0.0219 vs 0.0225) and NDCG (0.0267 vs 0.0270) are lower. The paper's explanation of 'instability of user modality preferences' is post hoc and not tested. The superiority claim should be restricted to the datasets and metrics where it holds, with significance tests or error bars.
[II-C, Node-Level Trade-off] The monotone relationship in Eq. (8), namely that larger |P_i,v − P_i,t| implies more difference amplification and smaller difference implies more alignment, is a stated hypothesis with no independent support. Because NLT is a central novel component, it needs a targeted diagnostic (e.g., comparing preference-difference magnitudes with per-node performance or embedding distances) rather than only the aggregate ablation in Table II.

minor comments (5)

[II-D, Eq. (10)] The text says 'L_t is omitted due to space constraints'; please provide the complete loss in an appendix for reproducibility.
[II-B] The phrase 'recommend system' should read 'recommendation systems'.
[II-C, Eq. (8)] The broadcast function fb is not defined precisely; it should specify how a per-node scalar is expanded to the feature dimension and whether the same weights are reused for user nodes.
[III-D, Fig. 2] The text refers to bar charts and a blue dotted line, but the experimental protocol for variants A and B is not fully specified; include dataset, hyperparameters, and how the fixed preference weights were chosen.
[III-B, Table I] No variance or significance measures are reported; given the small differences between top methods, error bars over repeated runs are necessary.

Circularity Check

0 steps flagged · score 0.0 of 10

No load-bearing circularity identified; the paper's sign inconsistency in Eq. (6) is a correctness flaw, not a circular reduction.

full rationale

The paper does not derive its performance gains from a fitted constant or self-citation chain. The node-level trade-off is an explicit hypothesis, with W_diff and W_cl defined from learned modality preferences rather than from the target metric. The homogeneous graph construction cites the authors' prior work, but that component is a borrowed building block, not the load-bearing justification for the claimed novelty, and the method is evaluated on external Amazon benchmarks against external baselines. Tuning hyperparameters on a validation set is normal model selection, not circular prediction. The most serious issue is in Section II-C, Eq. (6): L_diff = sigma_diff * || -|H*_{i,v} - H*_{i,t}| ⊙ W_diff ||_2 is minimized by shrinking the cross-modal difference, i.e., alignment, not by 'maximiz[ing] the differences between modality features.' That is an internal mathematical inconsistency, not a circularity, because it reverses rather than assumes the claimed mechanism. Accordingly, no circular step is identified.

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

The ledger shows that MDE introduces no new physical or conceptual entities. It relies on several hand-chosen hyperparameters, one unvalidated monotonicity assumption about node-level preferences, and a set of standard graph and recommendation modeling assumptions. The free parameters are all tuned or fixed on validation data, and their values are mostly unreported.

free parameters (4)

sigma_diff, sigma_cl, sigma_reg = grid-searched on validation set, values not reported
These weight the difference loss, contrastive loss, and regularization in Eq. 11; they are critical to the balance and are tuned per dataset.
K for user-user and item-item graphs = 10 (item) and 40 (user)
Top-K neighbors in homogeneous graph construction; fixed without sensitivity analysis.
GCN layers L = 2
Depth of LightGCN propagation in the heterogeneous graph; no ablation over this value.
Node-level preference matrix W_i = learned parameters, no reported distribution
Used in softmax fusion in Eq. 5 and drives the trade-off weights in Eq. 8; it is learned from user behavior data and is central to the method.

assumptions (4)

domain assumption Homogeneous user co-occurrence and item cosine-similarity graphs capture useful higher-order relations for multimodal recommendation
Section II-B constructs G_u and G_i from co-occurrence counts and feature cosine similarity; the effectiveness of these graphs is assumed and not independently justified.
domain assumption Pre-extracted visual and textual features from MMRec are fixed and sufficient; all gains come from graph propagation and fusion
Items are initialized with pre-trained features and no feature fine-tuning is performed, so the method inherits the quality and biases of those fixed features.
ad hoc to paper Larger |P_v - P_t| implies a node benefits more from difference amplification, while smaller |P_v - P_t| implies it benefits more from alignmen
The Node-Level Trade-off mechanism in Eq. 8 is built entirely on this stated hypothesis, with no independent evidence beyond the ablation table.
domain assumption BPR ranking loss with sampled negative items is an appropriate proxy for the recommendation objective
The final optimization in Section II-D uses BPR, a standard but still load-bearing modeling assumption for implicit feedback data.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MDE: Modality Discrimination Enhancement for Multi-modal Recommendation." pith.science (2026). https://pith.science/paper/N62WTK4R

@misc{pith2026250218481,
  author       = {Pith},
  title        = {Pith review of: MDE: Modality Discrimination Enhancement for Multi-modal Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N62WTK4R}},
  note         = {Machine review of arXiv:2502.18481}
}

read the original abstract

Multi-modal recommendation systems aim to enhance performance by integrating an item's content features across various modalities with user behavior data. Effective utilization of features from different modalities requires addressing two challenges: preserving semantic commonality across modalities (modality-shared) and capturing unique characteristics for each modality (modality-specific). Most existing approaches focus on aligning feature spaces across modalities, which helps represent modality-shared features. However, modality-specific distinctions are often neglected, especially when there are significant semantic variations between modalities. To address this, we propose a Modality Distinctiveness Enhancement (MDE) framework that prioritizes extracting modality-specific information to improve recommendation accuracy while maintaining shared features. MDE enhances differences across modalities through a novel multi-modal fusion module and introduces a node-level trade-off mechanism to balance cross-modal alignment and differentiation. Extensive experiments on three public datasets show that our approach significantly outperforms other state-of-the-art methods, demonstrating the effectiveness of jointly considering modality-shared and modality-specific features.

Figures

Figures reproduced from arXiv: 2502.18481 by the authors.

**Figure 1.** The framework of the proposed MDE. modality feature matrix for items as Ei m ∈ R|I|×dm under modality m, where dm denotes the feature dimension. The user historical interaction matrix is denoted as R ∈ {0, 1} |U |×|I|, in which the ru,i = 1 means the user u has interacted with item i, we construct a user-item interaction graph G = {U ∪ I, E}, where E = {(u, i) | u ∈ U, i ∈ I, Rui = 1} denotes the set of edges. The t… view at source ↗

**Figure 2.** Performance of different trade-off mechanism on Baby dataset [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 17 canonical work pages

[1]

Multimodal attentive learning for real-time explainable emotion recognition in conversations,

B. Arumugam, S. D. Bhattacharjee, and J. Yuan, “Multimodal attentive learning for real-time explainable emotion recognition in conversations,” in 2022 IEEE International Symposium on Circuits and Systems (IS- CAS). IEEE, 2022, pp. 1210–1214

work page 2022
[2]

Audio-visual cross- modal generation with multimodal variational generative model,

Z. Xu, T. Wang, D. Liu, D. Hu, H. Zeng, and J. Cao, “Audio-visual cross- modal generation with multimodal variational generative model,” in 2024 IEEE International Symposium on Circuits and Systems (ISCAS) . IEEE, 2024, pp. 1–5

work page 2024
[3]

Revisiting disentanglement and fusion on modality and context in conversational multimodal emotion recognition,

B. Li, H. Fei, L. Liao, Y . Zhao, C. Teng, T.-S. Chua, D. Ji, and F. Li, “Revisiting disentanglement and fusion on modality and context in conversational multimodal emotion recognition,” in Proceedings of the 31st ACM International Conference on Multimedia , 2023, pp. 5923– 5934

work page 2023
[4]

a3-fkg: Attentive attribute-aware fashion knowledge graph for outfit preference prediction,

H. Zhan, J. Lin, K. E. Ak, B. Shi, L.-Y . Duan, and A. C. Kot, “ a3-fkg: Attentive attribute-aware fashion knowledge graph for outfit preference prediction,” IEEE Transactions on Multimedia , vol. 24, pp. 819–831, 2021

work page 2021
[5]

Hierarchical multi-modal attention network for time-sync comment video recommendation,

W. Zhao, H. Wu, W. He, H. Bi, H. Wang, C. Zhu, T. Xu, and E. Chen, “Hierarchical multi-modal attention network for time-sync comment video recommendation,” IEEE Transactions on Circuits and Systems for Video Technology, 2023

work page 2023
[6]

Multimodal pretraining, adaptation, and generation for recommendation: A survey,

Q. Liu, J. Zhu, Y . Yang, Q. Dai, Z. Du, X.-M. Wu, Z. Zhao, R. Zhang, and Z. Dong, “Multimodal pretraining, adaptation, and generation for recommendation: A survey,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 6566– 6576

work page 2024
[7]

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

work page 2023
[8]

Lgmrec: Local and global graph learning for multimodal recommendation,

Z. Guo, J. Li, G. Li, C. Wang, S. Shi, and B. Ruan, “Lgmrec: Local and global graph learning for multimodal recommendation,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 8, 2024, pp. 8454–8462

work page 2024

Show all 25 references

[9]

A survey on contrastive self-supervised learning,

A. Jaiswal, A. R. Babu, M. Z. Zadeh, D. Banerjee, and F. Makedon, “A survey on contrastive self-supervised learning,” Technologies, vol. 9, no. 1, p. 2, 2020

2020
[10]

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
[11]

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 , vol. 25, pp. 5107–5116, 2022

2022
[12]

Drepmrec: A dual representation learning framework for multimodal recommendation,

K. Zhang, Y . Qin, R. Su, Y . Liu, J. Jin, W. Zhang, and Y . Yu, “Drepmrec: A dual representation learning framework for multimodal recommendation,” arXiv preprint arXiv:2404.11119 , 2024

2024 arXiv
[13]

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
[14]

Inductive representation learning on large graphs,

W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” Advances in neural information processing systems, vol. 30, 2017

2017
[15]

Semi-supervised classification with graph convolutional networks,

T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” in 5th International Conference on Learning Representations, ICLR, 2017

2017
[16]

Enhancing dyadic relations with homogeneous graphs for multimodal recommendation,

H. Zhou, X. Zhou, L. Zhang, and Z. Shen, “Enhancing dyadic relations with homogeneous graphs for multimodal recommendation,” in ECAI
[17]

Lightgcn: Simplifying and powering graph convolution network for recommenda- tion,

X. He, K. Deng, X. Wang, Y . Li, Y . Zhang, and M. Wang, “Lightgcn: Simplifying and powering graph convolution network for recommenda- tion,” in Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval , 2020, pp. 639– 648

2020
[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
[19]

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
[20]

Image-based recommendations on styles and substitutes,

J. McAuley, C. Targett, Q. Shi, and A. Van Den Hengel, “Image-based recommendations on styles and substitutes,” in Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval, 2015, pp. 43–52

2015
[21]

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
[22]

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
[23]

Adam: a method for stochastic optimization,

D. Kingma, “Adam: a method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014

2014 arXiv
[24]

Understanding the difficulty of training deep feedforward neural networks,

X. Glorot and Y . Bengio, “Understanding the difficulty of training deep feedforward neural networks,” in Proceedings of the thirteenth international conference on artificial intelligence and statistics . JMLR Workshop and Conference Proceedings, 2010, pp. 249–256

2010
[2023]

3123–3130

IOS Press, 2023, pp. 3123–3130

2023

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Multimodal attentive learning for real-time explainable emotion recognition in conversations,

B. Arumugam, S. D. Bhattacharjee, and J. Yuan, “Multimodal attentive learning for real-time explainable emotion recognition in conversations,” in 2022 IEEE International Symposium on Circuits and Systems (IS- CAS). IEEE, 2022, pp. 1210–1214

work page 2022

[2] [2]

Audio-visual cross- modal generation with multimodal variational generative model,

Z. Xu, T. Wang, D. Liu, D. Hu, H. Zeng, and J. Cao, “Audio-visual cross- modal generation with multimodal variational generative model,” in 2024 IEEE International Symposium on Circuits and Systems (ISCAS) . IEEE, 2024, pp. 1–5

work page 2024

[3] [3]

Revisiting disentanglement and fusion on modality and context in conversational multimodal emotion recognition,

B. Li, H. Fei, L. Liao, Y . Zhao, C. Teng, T.-S. Chua, D. Ji, and F. Li, “Revisiting disentanglement and fusion on modality and context in conversational multimodal emotion recognition,” in Proceedings of the 31st ACM International Conference on Multimedia , 2023, pp. 5923– 5934

work page 2023

[4] [4]

a3-fkg: Attentive attribute-aware fashion knowledge graph for outfit preference prediction,

H. Zhan, J. Lin, K. E. Ak, B. Shi, L.-Y . Duan, and A. C. Kot, “ a3-fkg: Attentive attribute-aware fashion knowledge graph for outfit preference prediction,” IEEE Transactions on Multimedia , vol. 24, pp. 819–831, 2021

work page 2021

[5] [5]

Hierarchical multi-modal attention network for time-sync comment video recommendation,

W. Zhao, H. Wu, W. He, H. Bi, H. Wang, C. Zhu, T. Xu, and E. Chen, “Hierarchical multi-modal attention network for time-sync comment video recommendation,” IEEE Transactions on Circuits and Systems for Video Technology, 2023

work page 2023

[6] [6]

Multimodal pretraining, adaptation, and generation for recommendation: A survey,

Q. Liu, J. Zhu, Y . Yang, Q. Dai, Z. Du, X.-M. Wu, Z. Zhao, R. Zhang, and Z. Dong, “Multimodal pretraining, adaptation, and generation for recommendation: A survey,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 6566– 6576

work page 2024

[7] [7]

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

work page 2023

[8] [8]

Lgmrec: Local and global graph learning for multimodal recommendation,

Z. Guo, J. Li, G. Li, C. Wang, S. Shi, and B. Ruan, “Lgmrec: Local and global graph learning for multimodal recommendation,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 8, 2024, pp. 8454–8462

work page 2024