Pith. sign in

REVIEW 4 major objections 8 minor 70 references

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

T0 review · 4 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that moving graph convolution from training to test time yields a multimodal recommender that beats 14 baselines while training in a fraction of the time.

desk verdict Genuinely new test-time-only GCN recipe for multimodal recommendation, but the paper's failure to specify whether the test graph includes held-out edges makes its headline gains potentially leaky. read the letter →

arxiv 2507.18489 v1 pith:XBVTMVPO submitted 2025-07-24 cs.IR

Jinfeng Xu , Zheyu Chen , Shuo Yang , Jinze Li , Edith C. H. Ngai This is my paper · ORCID
classification cs.IR
keywords multimodalrecommendationgraphconvolutionalnetworkstesting-phaseconvolutionBayesianpersonalizedrankingitem-itemmodalityisolationtrainingefficiencyrecommendersystems
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 argues that in multimodal recommendation, using graph convolution during training does more harm than good: the message-passing terms added to the ranking loss create positive and negative pairs that do not reflect real preferences, and neighbor aggregation pulls different modalities apart. It proposes FastMMRec, which trains a matrix-factorization-style model with a frozen item-item graph enhancement and only applies user-item graph convolution at test time. On Baby, Sports, and Clothing, this test-time-only model reports higher Recall@10/20 and NDCG@10/20 than fourteen baselines while cutting training time to roughly the level of plain matrix factorization. The paper stakes its contribution on the claim that the representational benefit of GCNs lies in aggregation at inference, not in gradient updates during training.

What carries the argument

The mechanism is a deliberate split of responsibilities between training and inference: training uses only BPR loss on multimodal embeddings plus an item-item graph enhancement, while testing applies LightGCN-style aggregation over the user-item graph. The analytical decomposition of one-layer graph convolution into Node-with-Node, Node-with-Neighbors, and Neighbors-with-Neighbors terms, and the corresponding split of the BPR loss into terms P1 through P5, carries the argument: terms P2 through P5 are claimed to be harmful because they assume interaction-neighborhood similarity that real data need not satisfy. The item-item graph is claimed safe for training because its neighbors are semantically similar items rather than users, so it avoids the same harmful-pair problem.

What would settle it

Run FastMMRec with user-item graph convolution included in training under the same hyperparameters and seeds on the same three datasets; if the trained-with-GCN version matches or beats the test-time-only version on Recall@20 and NDCG@20, the claim that training-time GCN inevitably creates net-harmful pairs is falsified.

Watch

Extended reading notes

Core claim

The central claim is that only adopting GCNs during the testing phase can enjoy the representational enhancement capabilities of GCNs and effectively circumvent the associated training challenges. Concretely, FastMMRec removes user-item graph convolution from training, keeps Bayesian Personalized Ranking on concatenated modality embeddings, enhances item representations with a frozen modality-aware item-item graph, and at test time applies several layers of LightGCN-style neighbor aggregation over the user-item interaction graph before scoring. The authors report that this combination beats all fourteen baselines on all three datasets, with absolute Recall@20 values of 0.1034, 0.1151, and 0.0992 on Baby, Sports, and Clothing, improvements over the best baseline between 4% and 7.5%, and per-epoch training times close to those of VBPR rather than graph-based models.

Load-bearing premise

The load-bearing premise is that the extra loss terms created by user-item graph convolution during training, namely P2 through P5, are net harmful; if those terms are helpful or neutral on the target data, removing them from training would not be the source of the reported gains.

Editorial extensions

If this is right

  • Per-epoch training cost drops to near matrix-factorization levels, so graph-based multimodal recommendation becomes practical where full user-item graph convolution in training was too expensive.
  • Because graph convolution at test time does not influence gradients, deeper aggregation of three or four layers keeps improving accuracy instead of triggering the over-smoothing seen when GCNs are trained.
  • Test-time graph convolution can be rewritten as recomputing user and item embeddings after training, keeping the reported inference overhead at roughly one percent over the fastest baseline.
  • Applying the same test-time-only strategy to existing graph-based models improves both accuracy and training speed, suggesting the principle extends beyond the proposed architecture.
  • Freezing a single fused item-item graph avoids the cubic graph-normalization cost of some baselines, reducing feature-mapping complexity to linear-in-items terms.

Reading between the lines

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

  • If the P2 through P5 decomposition is correct, the test-time-only aggregation idea may also transfer to non-multimodal collaborative filtering, where GCN training costs are equally dominated by neighbor sampling; that transfer is not tested in the paper.
  • The modality-isolation claim is measured by cosine similarity between visual and textual embeddings; a testable extension is to check whether pushing this similarity higher always helps recommendation, since some tasks may benefit from modality-specific specialization.
  • Because the test-time graph convolution is a fixed transformation, it could be applied on the fly to new users and items without retraining, suggesting an online-update path for streaming interactions that the paper does not evaluate.
  • The reported gains rest on three Amazon datasets, so a natural stress test is to vary dataset scale and domain where item-item graph semantics differ.
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 / 8 minor

Summary. The manuscript investigates the role of graph convolutional networks (GCNs) in multimodal recommendation. It argues that training with a user-item GCN inevitably creates harmful positive/negative pairs for BPR optimization and isolates different modalities, and it proposes FastMMRec, a framework that trains a lightweight model with a frozen item-item graph and applies user-item GCN only at inference. Experiments on Amazon Baby, Sports, and Clothing report that FastMMRec outperforms 14 baselines in Recall@K and NDCG@K while reducing training time and memory. The paper also includes ablations, sparsity studies, convergence curves, efficiency tables, and a t-SNE visualization.

Significance. If the empirical claims hold, the idea of deferring GCN message passing to the testing phase is an interesting and potentially practical contribution: it preserves part of the representational benefit of graph convolution while avoiding expensive training-time propagation. The strengths of the paper are its broad baseline comparison, consistent gains across three datasets, and additional analyses of sparsity, convergence, complexity, and memory. The claims are clearly stated and the main comparison is an honest A/B of training with versus without GCN. However, the theoretical framing is informal, and the central empirical result currently has a load-bearing ambiguity: the paper never specifies whether the test-time graph is built only from training edges. The contribution is therefore promising but not yet established.

major comments (4)
  1. [Section 4.2, Eqs. (18)-(19); Section 5.1.3] The test-time message passing in Eqs. (18)-(19) uses neighborhoods N(u) and N(i), but the manuscript never states that these neighborhoods are restricted to the training split. Section 5.1.3 describes an 8:1:1 random split for training/validation/test, yet the graph used at inference is not defined. If the adjacency matrix is built from the full interaction set, then for a test pair (u,i) the aggregated user embedding contains item i and the aggregated item embedding contains user u, so the score in Eq. (19) is mechanically inflated. This would invalidate the comparison in Table 3 and the MMGCN_train versus MMGCN_test results in Figure 1 and Table 9, because GCN baselines conventionally use a training-only adjacency matrix. I ask the authors to state explicitly which graph is used for test-time propagation, and if the full graph was used, to rerun all test-time evaluations with a training-only adjacency matrix and report whether the conclusions change.
  2. [Section 3.1, Eq. (7)] The claim that the BPR cross-terms P2-P5 are 'inevitably harmful' during optimization is not proven. Expanding the BPR loss into node-node, node-neighbor, and neighbor-neighbor terms does not by itself establish that those terms hurt optimization; their effect depends on the data distribution, the embedding geometry, and the training dynamics. The A/B experiments in Figure 1 and Table 9 are suggestive, but they do not identify the mechanism, and no variance estimates are provided. I recommend either providing a formal condition under which removing P2-P5 improves expected BPR loss, or weakening the 'inevitably' claim to an empirical observation.
  3. [Section 3.2, Eq. (8) and Table 1] The modality-isolation measure S, defined in Eq. (8) as the average cosine similarity between visual and textual embeddings, is never shown to be causally related to recommendation accuracy. The paper reports that MMGCN_test has higher S than MMGCN_train (Table 1) and that FastMMRec has the highest S among all models (Table 7), but this is a correlation. Higher S could also result from reduced training signal, modality collapse, or other confounding factors. To support challenge C2, an ablation that changes S while holding other components fixed, or another controlled experiment, is needed.
  4. [Section 4.1, Eq. (17)] The argument that item-item graph enhancement cannot create harmful positive/negative pairs is incomplete. Even though the neighbors are items rather than users, the negative term in Eq. (17) still decreases the similarity between the user and all neighbors of a sampled negative item. If any of those neighbors are actually relevant to the user, the gradient is harmful. The statement that semantic item neighbors 'do not encounter the inherent semantic discrepancies' does not address this negative-neighbor term. Please provide empirical evidence or a formal argument that these negative terms are benign.
minor comments (8)
  1. [Abstract and Section 1] The phrase 'GCNs inevitably create unhelpful and even harmful pairs' is stated as a fact in the Abstract and Section 1, but it is only an informal argument in Section 3. Please qualify the claim until the analysis is made rigorous.
  2. [Section 1] 'encompassing a amount of web multimedia content' should be 'encompassing a large amount of web multimedia content'.
  3. [Section 7] 'FastMMec' is a typo and should be 'FastMMRec'.
  4. [Table 8] The units of the reported times are ambiguous: values such as 7.17e-5 s appear to be per-user inference times, but this is not stated. Please define the unit explicitly.
  5. [Eq. (16)] The denominator k in the summation terms is not defined before first use; please clarify whether it is the number of kept neighbors or a normalization constant.
  6. [Table 3] The p-values are reported without specifying the statistical test or whether they are computed over users, over repeated runs, or with a paired test. Please describe the procedure.
  7. [Figure 1] The efficiency and effectiveness curves lack clearly labeled axes in the printed version; please add them.
  8. [Eq. (13)] The notation Con(alpha_m E_{um}) is unclear because alpha_m appears to be a scalar weight while the operation is concatenation; please clarify whether this is weighted concatenation or a different fusion operation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central test-phase-GCN claim rests on direct A/B comparisons and external baselines, not on fits or self-citation chains.

full rationale

The paper's derivation chain is not circular. Section 3 expands the LightGCN BPR loss into P1-P5 (Eq. 7), argues that P2-P5 can be unhelpful, and then tests this by A/B comparing MMGCN_train and MMGCN_test; the result is an empirical observation, not an identity or a fitted constant. FastMMRec's test-phase propagation (Eqs. 18-19) is evaluated against 14 external baselines (Table 3), with hyper-parameters selected on the validation split, so the reported gains are not forced by construction. Self-citations (e.g., [7, 8, 32, 36, 39]) appear only in related-work/context lists and do not supply the central result; no uniqueness theorem is imported. The 'harmful pairs' explanation is offered after the A/B result, but post-hoc mechanism attribution is not circular under the stated rules. The main scientific risk is a correctness/leakage ambiguity: Eqs. 18-19 do not specify whether N(u) and N(i) are computed from the training-only or the full graph, which would be label leakage rather than circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No new physical entities or mediators are postulated. The framework rests on several untested modeling assumptions listed above; the main free parameters are validation-tuned hyperparameters controlling graph depth and sparsity.

free parameters (4)
  • Test-time GCN depth L = 3 or 4 across datasets, tuned per Fig 6
    Grid-searched over {1,2,3,4}; the number of layers at test time is a central component of the method.
  • Item-item graph hops L_i = 2 (Baby), 1 (Sports), 1 (Clothing) per Fig 7
    Grid-searched over {1,2,3}; controls how many neighbor hops are aggregated during training.
  • Top-k item neighbors k = 10 (Baby), 10 (Sports), 5 (Clothing) per Fig 7
    Grid-searched over {5,10,15,20}; controls the sparsity of the frozen item-item graph.
  • Learning rate and L2 weight = 1e-4 and 1e-3, fixed in all experiments
    Empirically fixed; part of the reported configuration, but not tuned per dataset.
assumptions (5)
  • ad hoc to paper The BPR loss cross-terms P2-P5 in Eq (7) are harmful during training and should be avoided.
    The paper labels these terms 'unhelpful or harmful' based on real-world observations, not a mathematical proof; this motivates removing GCNs from training.
  • ad hoc to paper Higher cosine similarity S between visual and textual embeddings indicates better cross-modal alignment and less modality isolation.
    Used to validate challenge C2 via Table 1; no causal link to recommendation accuracy is established.
  • ad hoc to paper Item-item graph neighbors, being semantically similar items, do not introduce harmful positive/negative pairs during BPR optimization.
    Underpins the training-phase design; ignores false-negative risks when a user has not interacted with a semantically similar item.
  • domain assumption Pre-trained visual (4096-d) and text (384-d) features are adequate modality representations.
    All embeddings derive from these frozen encoders; the framework inherits their limitations.
  • domain assumption The transductive user-item graph at test time is the same as the training graph and fixed.
    Test-time GCN in Eq (18) aggregates over the training interaction graph; new interactions are not supported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Best is Yet to Come: Graph Convolution in the Testing Phase for Multimodal Recommendation." pith.science (2026). https://pith.science/paper/XBVTMVPO

@misc{pith2026250718489,
  author       = {Pith},
  title        = {Pith review of: The Best is Yet to Come: Graph Convolution in the Testing Phase for Multimodal Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XBVTMVPO}},
  note         = {Machine review of arXiv:2507.18489}
}
read the original abstract

The efficiency and scalability of graph convolution networks (GCNs) in training recommender systems remain critical challenges, hindering their practical deployment in real-world scenarios. In the multimodal recommendation (MMRec) field, training GCNs requires more expensive time and space costs and exacerbates the gap between different modalities, resulting in sub-optimal recommendation accuracy. This paper critically points out the inherent challenges associated with adopting GCNs during the training phase in MMRec, revealing that GCNs inevitably create unhelpful and even harmful pairs during model optimization and isolate different modalities. To this end, we propose FastMMRec, a highly efficient multimodal recommendation framework that deploys graph convolutions exclusively during the testing phase, bypassing their use in training. We demonstrate that adopting GCNs solely in the testing phase significantly improves the model's efficiency and scalability while alleviating the modality isolation problem often caused by using GCNs during the training phase. We conduct extensive experiments on three public datasets, consistently demonstrating the performance superiority of FastMMRec over competitive baselines while achieving efficiency and scalability.

Figures

Figures reproduced from arXiv: 2507.18489 by the authors.

Figure 1
Figure 1. Effectiveness and Efficiency study. In addition to the challenge C1 posed by GCNs within each modality, we further examine how GCNs affect the similarity be￾tween different modalities. To analyze this, we revisit Equations 6 and 7. For MF in Equation 6, the model learns specific weights for each user’s modalities by directly optimizing the nodes within each modality. Conversely, for GCN in Equation 7, the aggregatio… view at source ↗
Figure 2
Figure 2. The overall framework of the proposed multimodal recommendation model (FastMMRec). [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Sparsity degree analysis on three datasets. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Convergence study in terms of Recall@20. 5.5.2 Convergence [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 7
Figure 7. Figure 7: Effect of item-item graph hops 𝐿𝑖 and 𝑘 value. 5.7 Hyper-parameter Study We examine the sensitivity of several important hyper-parameters of FastMMRec across different datasets. • GCN layers 𝐿: We first investigate the impact of GCN depth by varying the number of messa…
Figure 5
Figure 5. Figure 5: Distribution of visual and textual representations [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Effect of GCN layers 𝐿. 1 2 3 Item-item graph hops Li 5 10 15 20 v alu e k 0.1017 0.1025 0.1013 0.1029 0.1034 0.1023 0.0997 0.1018 0.1002 0.0994 0.1010 0.0999 Recall@20 0.092 0.094 0.096 0.098 0.100 0.102 (a) Baby 1 2 3 Item-item graph hops Li 5 10 15 20 v alu e k 0.11…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

70 extracted references · 42 canonical work pages

  1. [1]

    Feiyu Chen, Junjie Wang, Yinwei Wei, Hai-Tao Zheng, and Jie Shao. 2022. Break- ing isolation: Multimodal graph fusion for multimedia recommendation by edge- wise modulation. In Proceedings of the 30th ACM International Conference on Multimedia. 385–394

    work page 2022

  2. [2]

    Hao Chen, Yuanchen Bei, Qijie Shen, Yue Xu, Sheng Zhou, Wenbing Huang, Feiran Huang, Senzhang Wang, and Xiao Huang. 2024. Macro graph neural networks for online billion-scale recommender systems. In Proceedings of the ACM on Web Conference 2024. 3598–3608

    work page 2024

  3. [3]

    Jie Chen, Tengfei Ma, and Cao Xiao. 2018. FastGCN: Fast Learning with Graph Convolutional Networks via Importance Sampling. In International Conference on Learning Representations

    2018

  4. [4]

    Jingyuan Chen, Hanwang Zhang, Xiangnan He, Liqiang Nie, Wei Liu, and Tat- Seng Chua. 2017. Attentive 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 Retrieval . 335–344

    work page 2017

  5. [5]

    Jianfei Chen, Jun Zhu, and Le Song. 2018. Stochastic Training of Graph Convolu- tional Networks with Variance Reduction. InInternational Conference on Machine Learning. PMLR, 942–950

    work page 2018

  6. [6]

    Xu Chen, Hanxiong Chen, Hongteng Xu, Yongfeng Zhang, Yixin Cao, Zheng Qin, and Hongyuan Zha. 2019. Personalized fashion recommendation with visual explanations based on multimodal attention network: Towards visually explainable recommendation. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieva...

    work page 2019

  7. [7]

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

    work page 2025

  8. [8]

    Zheyu Chen, Jinfeng Xu, Yutong Wei, and Ziyue Peng. 2025. Squeeze and Ex- citation: A Weighted Graph Contrastive Learning for Collaborative Filtering. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval . 2769–2773

    work page 2025

Show all 70 references
  1. [9]

    Weilin Cong, Rana Forsati, Mahmut Kandemir, and Mehrdad Mahdavi. 2020. Minimal variance sampling with provable guarantees for fast training of graph neural networks. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining . 1393–1403

    2020

  2. [10]

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

    2010

  3. [11]

    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

  4. [12]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. Advances in neural information processing systems 30 (2017)

    2017

  5. [13]

    Xiaotian Han, Tong Zhao, Yozen Liu, Xia Hu, and Neil Shah. 2023. MLPInit: Embarrassingly Simple GNN Training Acceleration with MLP Initialization. In The Eleventh International Conference on Learning Representations

    2023

  6. [14]

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

    2016

  7. [15]

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

    2020

  8. [16]

    Yang Hu, Haoxuan You, Zhecan Wang, Zhicheng Wang, Erjin Zhou, and Yue Gao. 2021. Graph-mlp: Node classification without message passing in graph. arXiv preprint arXiv:2106.04051 (2021)

    2021 arXiv

  9. [17]

    Yangqin Jiang, Lianghao Xia, Wei Wei, Da Luo, Kangyi Lin, and Chao Huang

  10. [18]

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

    2014 arXiv

  11. [19]

    Meng Liu, Hongyang Gao, and Shuiwang Ji. 2020. Towards deeper graph neural networks. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining . 338–348

    2020

  12. [20]

    Shang Liu, Zhenzhong Chen, Hongyi Liu, and Xinghai Hu. 2019. User-video co- attention network for personalized micro-video recommendation. In The world wide web conference. 3020–3026

    2019

  13. [21]

    Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel

  14. [22]

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

  15. [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 Multimedia (2022)

    2022

  16. [24]

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

    2008

  17. [25]

    Lin Wang, Wenqi Fan, Jiatong Li, Yao Ma, and Qing Li. 2024. Fast graph conden- sation with structure-based neural tangent kernel. In Proceedings of the ACM on Web Conference 2024. 4439–4448

    2024

  18. [26]

    Qifan Wang, Yinwei Wei, Jianhua Yin, Jianlong Wu, Xuemeng Song, and Liqiang Nie. 2021. Dualgnn: Dual graph neural network for multimedia recommendation. IEEE Transactions on Multimedia (2021)

    2021

  19. [27]

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

    2023

  20. [28]

    Wei Wei, Xubin Ren, Jiabin Tang, Qinyong Wang, Lixin Su, Suqi Cheng, Jun- feng Wang, Dawei Yin, and Chao Huang. 2024. Llmrec: Large language models with graph augmentation for recommendation. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining...

    2024

  21. [29]

    Wei Wei, Jiabin Tang, Lianghao Xia, Yangqin Jiang, and Chao Huang. 2024. Promptmm: Multi-modal knowledge distillation for recommendation with prompt-tuning. In Proceedings of the ACM on Web Conference 2024 . 3217–3228

    2024

  22. [30]

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

    2020

  23. [31]

    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. In Proceedings of the 27th ACM international conference on multimedia. 1437–1445

    2019

  24. [32]

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

    2025 arXiv

  25. [33]

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

  26. [34]

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

    2024 arXiv

  27. [35]

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

    2024

  28. [36]

    Jinfeng Xu, Zheyu Chen, Wei Wang, Xiping Hu, Sang-Wook Kim, and Edith CH Ngai. 2025. COHESION: Composite Graph Convolutional Network with Dual- Stage Fusion for Multimodal Recommendation. In Proceedings of the 48th In- ternational ACM SIGIR Conference on Research and Developme...

    2025

  29. [37]

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

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

  30. [38]

    Jinfeng Xu, Zheyu Chen, Shuo Yang, Jinze Li, Hewei Wang, Wei Wang, Xiping Hu, and Edith Ngai. 2025. NLGCL: Naturally Existing Neighbor Layers Graph Contrastive Learning for Recommendation. arXiv preprint arXiv:2507.07522 (2025)

    2025 arXiv

  31. [39]

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

    2025 arXiv

  32. [40]

    Liangwei Yang, Zhiwei Liu, Chen Wang, Mingdai Yang, Xiaolong Liu, Jing Ma, and Philip S Yu. 2023. Graph-based alignment and uniformity for recommendation. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management. 4395–4399

    2023

  33. [41]

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

  34. [42]

    Zixuan Yi and Iadh Ounis. 2024. A unified graph transformer for overcom- ing isolations in multi-modal recommendation. In Proceedings of the 18th ACM Conference on Recommender Systems . 518–527

    2024

  35. [43]

    Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L Hamilton, and Jure Leskovec. 2018. Graph convolutional neural networks for web-scale recommender systems. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining . 974–983

    2018

  36. [44]

    Junliang Yu, Hongzhi Yin, Xin Xia, Tong Chen, Lizhen Cui, and Quoc Viet Hung Nguyen. 2022. Are graph augmentations necessary? simple graph contrastive learning for recommendation. In Proceedings of the 45th international ACM SIGIR MM ’25, October 27–31, 2025, Dublin, Ireland X...

    2022

  37. [45]

    Penghang Yu, Zhiyi Tan, Guanming Lu, and Bing-Kun Bao. 2023. Multi-view graph convolutional network for multimedia recommendation. In Proceedings of the 31st ACM International Conference on Multimedia . 6576–6585

    2023

  38. [46]

    Wenjie Yang, Shengzhong Zhang, Jiaxing Guo, and Zengfeng Huang. 2024. Your Graph Recommender is Provably a Single-view Graph Contrastive Learning. arXiv preprint arXiv:2407.17723 (2024)

    2024 arXiv

  39. [47]

    Jiahao Zhang, Rui Xue, Wenqi Fan, Xin Xu, Qing Li, Jian Pei, and Xiaorui Liu

  40. [48]

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

  41. [49]

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

  42. [50]

    Shichang Zhang, Yozen Liu, Yizhou Sun, and Neil Shah. 2022. Graph-less Neural Networks: Teaching Old MLPs New Tricks Via Distillation. In International Conference on Learning Representations

    2022

  43. [51]

    Hanqing Zeng, Hongkuan Zhou, Ajitesh Srivastava, Rajgopal Kannan, and Viktor Prasanna. 2019. GraphSAINT: Graph Sampling Based Inductive Learning Method. In International Conference on Learning Representations

    2019

  44. [52]

    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

  45. [53]

    In Proceedings of the ACM on Web Conference 2024

    Linear-Time Graph Neural Networks for Scalable Recommendations. In Proceedings of the ACM on Web Conference 2024 . 3533–3544

    2024

  46. [54]

    Xin Zhou. 2023. MMRec: Simplifying Multimodal Recommendation. arXiv preprint arXiv:2302.03497 (2023)

    2023 arXiv

  47. [55]

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

    2023

  48. [56]

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

    2023

  49. [57]

    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. In Proceedings of the ACM Web Conference 2023 . 845–854

    2023

  50. [58]

    Difan Zou, Ziniu Hu, Yewen Wang, Song Jiang, Yizhou Sun, and Quanquan Gu

  51. [59]

    Lingxiao Zhao and Leman Akoglu. 2020. PairNorm: Tackling Oversmoothing in GNNs. In International Conference on Learning Representations

    2020

  52. [61]

    Hongyu Zhou, Xin Zhou, Lingzi Zhang, and Zhiqi Shen. 2023. Enhancing dyadic relations with homogeneous graphs for multimodal recommendation. In ECAI

    2023

  53. [69]

    • MMGCN [31] applies a GCN for each modality to learn modality- specific features and then integrates all user-predicted ratings across modalities to produce the final rating

    Multimodal recommendation models: • VBPR [14] combines visual and textual features with ID embed- dings as side information for each item, effectively achieving multimodal matrix factorization. • MMGCN [31] applies a GCN for each modality to learn modality- specific features a...

    2025

  54. [70]

    stabilizes these representations by freezing the item semantic graph. In a novel approach, MMSSL [27] and MICRO [49] employ contrastive self-supervised learning to align modalities and collab- orative signals to enhance recommendation Additionally, BM3 [57] and PromptMM [29] i...

  55. [2012]

    arXiv preprint arXiv:1205.2618 (2012)

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

    2012 arXiv

  56. [2015]

    In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval

    Image-based recommendations on styles and substitutes. In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval. 43–52

  57. [2019]

    Advances in neural information processing systems 32 (2019)

    Layer-dependent importance sampling for training deep and large graph convolutional networks. Advances in neural information processing systems 32 (2019). FastMMRec MM ’25, October 27–31, 2025, Dublin, Ireland Table 7: Similarity𝑆 between visual and textual embeddings. Models ...

    2019

  58. [2021]

    In Proceedings of the 29th ACM International Conference on Multimedia

    Mining latent structures for multimedia recommendation. In Proceedings of the 29th ACM International Conference on Multimedia . 3872–3880

  59. [2022]

    IEEE Transactions on Knowledge and Data Engineering 35, 9 (2022), 9154–9167

    Latent structure mining with contrastive modality fusion for multimedia recommendation. IEEE Transactions on Knowledge and Data Engineering 35, 9 (2022), 9154–9167

    2022

  60. [2023]

    IOS Press, 3123–3130

  61. [2024]

    DiffMM: Multi-Modal Diffusion Model for Recommendation. (2024)

    2024

  62. [2025]

    In Proceedings of the AAAI Conference on Artificial Intelligence , Vol

    Mentor: multi-level self-supervised learning for multimodal recommen- dation. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 39. 12908–12917

Pith tools

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