Pith. sign in

REVIEW 5 major objections 6 minor 47 references

One carefully refined item-item graph, reused three ways, beats specialized multimodal recommenders while using less compute.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-31 10:44 UTC pith:LNMAGFUM

load-bearing objection Solid systems packaging of one cleaned item–item graph reused three ways; empirics are the real load-bearing part, and the NCER hub story is softer than the theorems claim. the 5 major comments →

arxiv 2607.24607 v1 pith:LNMAGFUM submitted 2026-07-27 cs.IR

One Graph, Multiple Gains: Single High-Quality Item-Item Graph for Multimodal Recommendation

Jinfeng Xu , Zheyu Chen , Ziyue Peng , Shuo Yang , Jinze Li , Zewei Liu , Shujie Li , Yipeng Du
show 1 more author
Edith C. H. Ngai
This is my paper
classification cs.IR
keywords multimodal recommendationitem-item graphneighborhood consistencytriadic closureresidual gateBPR augmentationcold-startgraph denoising
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Multimodal recommenders already build item-item graphs from images, text, and behavior, but those graphs are usually noisy and used only to push item embeddings around. This paper argues that the real gains come from building one high-quality item-item graph once, then deliberately reusing it at three different stages of the pipeline. The graph is formed by fusing content similarity with co-occurrence, then reweighted by how many neighbors two items share (triadic closure), so reliable edges are amplified and isolated edges are downweighted. That same graph then (1) feeds a per-item gate that decides how much semantic neighborhood signal each item should absorb, (2) expands the user-item interaction graph with virtual edges through high-confidence neighbors, and (3) supplies discounted soft positives for ranking loss. Theory claims the reweighting lowers spectral noise, the gate does not collapse, and the soft positives tighten the generalization bound. On four datasets the resulting system outperforms strong baselines, runs faster, uses less GPU memory, and helps most when interactions are sparse or items are cold-start.

Core claim

A single item-item graph, refined once by neighborhood-consistency reweighting and then systematically reused for representation gating, interaction-graph expansion, and BPR augmentation, is enough to outperform state-of-the-art multimodal recommenders while cutting training time and memory, with the largest gains under cold-start and sparse conditions.

What carries the argument

Neighborhood Consistency Edge Reweighting (NCER): edges in a fused semantic-plus-co-occurrence item-item graph are multiplied by (1 + κ · shared-neighbor fraction), which the authors prove reduces the Frobenius noise-to-signal ratio when true edges have higher triadic overlap than noise edges; that refined graph then drives Residual II Gate (RIG) and II-Neighbor BPR Augmentation (INA).

Load-bearing premise

True semantic edges systematically share more neighbors than noisy edges, so boosting high-overlap connections actually cleans the graph rather than amplifying hubs or cross-community accidents.

What would settle it

On a dataset where noise edges have equal or higher shared-neighbor scores than true edges, NCER should stop improving (or hurt) Recall/NDCG relative to the unreweighted fused graph, and the cold-start and sparsity gains should shrink or vanish.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Precomputing one refined item-item graph can replace multiple ad-hoc modality graphs and still improve ranking quality.
  • Per-item adaptive gates can safely absorb semantic neighbors without forcing every item to take the same propagation strength.
  • Top neighbors of observed positives can be used as discounted soft labels without conflicting with standard BPR.
  • Cold-start and low-activity users benefit most when virtual edges and soft positives are drawn from a noise-reduced item graph.
  • Training cost can fall because the expensive graph work is done once offline and cached.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same triadic-closure filter could be tried on pure collaborative item-item graphs in non-multimodal settings where content features are absent.
  • If hub items still dominate shared-neighbor counts, an extra degree-normalized consistency score might be needed before NCER is safe at web scale.
  • RIG’s learned gates could be inspected post-hoc as a diagnostic of which items have trustworthy multimodal neighborhoods.
  • The build-once, reuse-everywhere pattern may transfer to session or sequential recommenders that already maintain item affinity tables.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes IIMRec, a multimodal recommendation framework built around a single item-item graph constructed once in preprocessing (fusing semantic KNN graphs over spatial/DCT features with a co-occurrence graph), refined by Neighborhood Consistency Edge Reweighting (NCER), and reused at three pipeline stages: a Residual II Gate (RIG) for item-item propagation, content-guided expansion of the user-item interaction graph, and II-Neighbor BPR Augmentation (INA) that treats top semantic neighbors of positives as discounted soft positives. The paper provides three theorems (NCER noise-ratio reduction, RIG non-degeneracy, INA PAC-Bayes generalization) and an extensive empirical study on four datasets (Baby, Sports, Clothing, TikTok) against a strong battery of recent multimodal baselines, with ablations, sparsity/cold-start splits, an efficiency table, and hyperparameter sweeps. IIMRec wins on essentially every metric, with the largest margins under cold-start and sparse conditions.

Significance. If the results hold, the contribution is a well-executed, practically useful recipe: the "build once, leverage everywhere" framing is a clean organizational idea, the empirical evaluation is comparatively strong for this venue class (four datasets, recent strong baselines including SMORE, MENTOR, COHESION, HPMRec, significance tests, component ablations, sparsity/cold-start analyses, and an efficiency study showing IIMRec is both faster and lighter than the baselines it beats), and code is promised. The efficiency gains (Table 4: roughly 2-6x faster per epoch and substantially less GPU memory than the strongest competitors) are, if reproducible, as consequential as the accuracy gains. The cold-start results in Table 3 are the most persuasive evidence that the semantic graph is doing real work rather than merely amplifying popularity. However, the theoretical scaffolding is weaker than the empirical core: the central noise-suppression premise of NCER is structurally contradicted by the hub-noise model the paper itself invokes, Theorem 5.3 is close to vacuous, and Theorem 5.5 imports assumptions its proof does not justify. These do not necessarily undermine the empirical claims, but,

major comments (5)
  1. [§5.1, Theorem 5.2 and Eq. (7)-(8)] The load-bearing premise of NCER is contradicted by the noise model the paper itself names. §1/§2 motivate NCER as a remedy for 'hub effects and feature contamination,' and the co-occurrence graph C = R^T R (Eq. 5) is explicitly popularity-driven. But hub noise is precisely the failure mode of Assumption 5.1: an edge incident on a hub item shares many common neighbors by construction (the hub's degree inflates (BB^T)_ij in Eq. 7), so hub-induced noise edges receive high c_ij and are amplified by the (1 + kappa*c_ij) factor in Eq. 8, not suppressed. Moreover, c_ij is computed from B, the binarized fused graph that already contains the noise: the measurement of 'structural reliability' and the quantity being measured are the same contaminated signal, so Theorem 5.2's condition c_bar_S* > c_bar_N cannot be verified from the observable quantities the proof uses. This needs either (a) an empi
  2. [Theorem 5.2 vs. Appendix A, Eqs. (17) and (26)] There is a statement-level mismatch between the main-text theorem and its proof. Eq. (17) bounds ||N_hat||_F / ||S_hat||_F (ratio against the full reweighted observed graph), while the appendix derives Eq. (26) in terms of ||N_hat||_F / ||S*_hat||_F (ratio against the clean component) and only gestures at the conversion via an epsilon/||S*||_F correction in the final sentence. Additionally, Eq. (25) claims ||S* o W||_F^2 >= ||S*||_F^2 (1 + kappa c_bar_S*)^2, but W_ii-type entries with c < c_bar pull the norm the other way; the inequality as written only holds with the corresponding O(kappa^2 Var) term that the noise side carries. The proof should be rewritten so the claimed bound and the derived bound are the same object, with the variance terms handled symmetrically.
  3. [§5.2, Theorem 5.3 and Appendix B] The non-degeneracy result is close to vacuous as stated: since the gate is a sigmoid output, g_i in (0,1) holds by construction for every item, regardless of the gradient analysis. The substantive question — whether gates saturate arbitrarily close to 0 or 1, defeating per-item adaptivity — is not addressed. Worse, the stationary-point argument in Appendix B Case 1 does not establish an interior equilibrium: the positive-side and negative-side gradient contributions (Eqs. 30 and 33) have opposite signs, and the claim that 'the balance is achieved at an interior point because the BPR weights depend on the margin' is an assertion, not an argument; nothing rules out the stationary point lying at the effective boundary. Recommendation: either prove a real statement (e.g., gradients bounded away from zero in a neighborhood of saturation under stated conditions) or downgrade Theorem 5.3 to qua
  4. [§5.3, Theorem 5.5 and Appendix C] The PAC-Bayes argument has two unjustified steps. First, the augmented triplets (u, j, n') are deterministic functions of the observed interactions (j is drawn from N_II(p) where (u,p) is an observed edge), so the n + delta_eff*n_aug samples are not i.i.d. and the standard bound (Eq. 35) does not transfer to the augmented denominator without an explicit dependence correction. Second, the effective sample size delta_eff = delta^2 * q is imported by analogy to noisy-label learning [11] without derivation; the claim that discount delta reduces 'both the gradient magnitude (linear in delta) and the effective signal strength (linear in delta*q)' and hence contributes delta^2*q samples is a heuristic, not a consequence of the cited framework. Also note a symbol collision: delta is simultaneously the INA discount factor (Eq. 13) and the PAC-Bayes confidence parameter (log(1/delta) in Eq. 19), a
  5. [§6.3, Figure 2 (ablation design)] The ablation does not isolate the claimed mechanism. w/o NCER shows NCER helps, but any reweighting that amplifies community-dense/popular edges would also help on these benchmarks — popularity amplification is itself a strong prior, and it is the opposite of the claimed 'noise suppression' mechanism. A cheap, decisive control exists: compare NCER against (i) degree/popularity-proportional reweighting (1 + kappa * deg-normalized popularity of i,j) and (ii) random reweighting with the same marginal distribution of multipliers. If NCER beats (i), the triadic-closure mechanism is supported; if it merely matches, the contribution is a well-tuned popularity prior and the paper's framing should be revised. Given that the paper's central pitch is 'high-quality graph' rather than 'better-weighted graph,' this control is load-bearing for the stated contribution.
minor comments (6)
  1. [Figure 2] Figure 2's variant labels are corrupted in the rendered PDF ('w/o /glyph1197CER', 'w/o I/glyph1197A'), making the two most important ablations (NCER, INA) hard to identify; please fix the font embedding.
  2. [§4.7, §6.5, Table 4] The 'zero overhead during training' claim should be complemented by reporting the one-time preprocessing cost (multi-modality KNN construction, DCT, BB^T for NCER, neighbor table). Table 4 is per-epoch only; a wall-clock total (preprocessing + training to convergence) would make the efficiency claim airtight.
  3. [Multiple equations] Typesetting artifacts throughout: 'Í' appears in place of a summation symbol in Eqs. (2), (10), (14), (15); Eq. (10)'s left-hand side is labeled e_u where it should be the fused user embedding expression; missing period after Eq. (6).
  4. [§4.3, §4.6, §5.3] Notation collision: delta denotes the INA discount (Eq. 13), the PAC-Bayes confidence (Theorem 5.5), and appears in delta_eff = delta^2*q using the INA meaning — please rename one. Similarly, k is used both as the KNN neighbor count and as the normalization constant in Eq. (7); since row-wise top-k gives |N(i)| = k this is consistent, but a sentence noting that (BB^T)_ij <= k would preempt confusion.
  5. [§6.4] Cold-start protocol (§6.4) cites 'prior setting [22]' (the authors' own work); a brief description of how cold-start items receive modality features and UI-graph edges (they have no interactions, so CG-UI virtual edges flow one way) would help reproducibility.
  6. [References] Reference [14] (BPR) is cited as arXiv:1205.2618; the canonical citation is the UAI 2009 paper.

Circularity Check

0 steps flagged

No significant circularity: theorems are conditional inequalities from stated assumptions; main claims are held-out empirical comparisons.

full rationale

IIMRec’s load-bearing claims are (i) empirical gains vs external baselines on held-out Recall/NDCG (Tables 2–4, Fig. 3) and (ii) three conditional theory results. Theorem 5.2 states that NCER reduces the Frobenius noise-to-signal ratio when clean edges have higher average neighborhood consistency than noise edges (Assumption 5.1); the proof multiplies edge weights by (1+κ c_ij) and compares ratios under that premise—it does not define the noise ratio as the reweighting, nor fit c from the target metric. Theorem 5.3 derives the BPR gradient w.r.t. the RIG gate and argues interior stationary points under regularity; the gate is learned, not set equal to the claimed non-degeneracy. Theorem 5.5 is a standard PAC-Bayes bound with an effective sample-size term δ_eff·n_aug that is tighter only when soft positives have positive predictive value—again conditional, not tautological. Hyperparameters (k, κ, γ, δ, λ) are validation-tuned and evaluated on test rankings, which is ordinary supervised practice, not fitted-input-called-prediction of the reported metrics. Author self-citations appear as related work and baselines (COHESION, MENTOR, NLGCL, etc.) but are not invoked as uniqueness theorems that force the method; central comparisons include independent SOTA systems. Concerns that hub noise can violate Assumption 5.1 (triadic closure amplifying hubs) are assumption-validity / correctness issues, not derivation circularity. No step reduces a claimed prediction to its inputs by construction.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 3 invented entities

The load-bearing story depends on standard recsys losses and GCN backbones, plus domain assumptions that neighborhood overlap marks clean item-item edges and that semantic neighbors of positives are useful soft labels. Several scalar knobs are grid-searched. The named modules are engineered mechanisms, not new physical entities; their value is justified mainly by ablations and the community-structure noise model.

free parameters (7)
  • k (Top-k neighbors for all KNN / INA tables) = 10 (typical optimum)
    Shared neighbor budget for semantic, co-occurrence, filtered UI expansion, and soft positives; grid-searched in {5,10,15,20}, peak reported at 10.
  • κ (NCER reweighting strength) = 0.4 (0.6 on TikTok)
    Controls how strongly shared-neighbor consistency amplifies edges; searched in {0.2,0.4,0.6,0.8}.
  • γ (content-guided UI blend) = 0.2–0.4 preferred
    Mixes original interactions with virtual edges R S_filtered; searched in {0.2,0.4,0.6,0.8}.
  • δ (INA soft-positive discount) = stable in [0.2,0.6]
    Scales soft BPR terms; searched in {0.2,0.4,0.6,0.8}.
  • λ_INA and λ_CL (loss weights) = λ_INA≈0.6; λ_CL≈1e-4
    Trade off primary BPR against INA and InfoNCE alignment; grid-searched.
  • ρ (DCT low/high frequency split) = 0.5
    Fixed band split for spectral KNN graphs.
  • temperatures τ, τ_cl = 0.2
    Fixed softmax/InfoNCE temperatures to reduce tuning.
axioms (6)
  • domain assumption Observed item-item similarity decomposes as clean community-structured S* plus bounded noise N; clean edges have higher neighborhood overlap than noise edges (Assumption 5.1).
    Required for Theorem 5.2’s strict noise-to-signal reduction under NCER.
  • domain assumption Triadic closure / shared-neighbor count is a valid structural reliability signal on item-item affinity graphs built from multimodal features and co-occurrence.
    Motivates NCER Eq. (7)–(8); imported from social-network theory (Easley & Kleinberg) without independent item-graph proof.
  • domain assumption Top neighbors of an interacted item are more likely user-preferred than random non-interacted items, so discounted soft BPR pushes with positive effective sample size δ_eff (INA / Theorem 5.5).
    Supported by citation to prior modality-neighbor overlap work; quality factor q is not measured on these datasets.
  • domain assumption LightGCN-style linear propagation on (expanded) UI graphs and BPR pairwise ranking are appropriate primary objectives for implicit multimodal recommendation.
    Standard recsys backbone choices in Section 4.4 and 4.7.
  • standard math Frobenius-norm noise-to-signal ratio and PAC-Bayes bounds with effective noisy-label sample size are meaningful proxies for recommendation generalization here.
    Proof frameworks in Appendices A and C; standard inequalities plus learning-with-noisy-labels style effective sample size.
  • ad hoc to paper Gate MLP outputs remain interior (0,1) under bounded embeddings and non-degenerate initialization so RIG cannot hard-collapse (Theorem 5.3).
    Regularity conditions stated for non-degeneracy; not verified beyond gradient sign arguments.
invented entities (3)
  • NCER (Neighborhood Consistency Edge Reweighting) no independent evidence
    purpose: Reweight fused item-item edges by normalized shared-neighbor counts to suppress spurious similarity/co-occurrence links.
    Core preprocessing mechanism; independent evidence is indirect (ablations + spectral inequality under Assumption 5.1), not an external measurable object.
  • RIG (Residual II Gate) no independent evidence
    purpose: Per-item learned scalar gate controlling how much II-propagated semantics enter the final item embedding.
    Addresses uniform residual aggregation; justified by end-to-end BPR gradients and ablation drops when removed.
  • INA (II-Neighbor BPR Augmentation) no independent evidence
    purpose: Treat top II neighbors of positives as discounted soft positives to densify ranking supervision.
    Optimization-stage reuse of the same graph; depends on neighbor predictive value q not directly estimated.

pith-pipeline@v1.2.0-grok45-kimik3 · 26530 in / 4090 out tokens · 62003 ms · 2026-07-31T10:44:08.117201+00:00 · methodology

0 comments
read the original abstract

Multimodal recommendation leverages item multimodal features alongside collaborative signals to capture user preferences. While item-item graphs have become a key building block in advanced models, existing methods typically construct them with noisy similarity edges and limit their role to a single function of item-item representation propagation, leaving substantial potential untapped. In this paper, we propose IIMRec, a framework that constructs a single high-quality item-item graph during preprocessing and systematically reuses it across three stages of the recommendation pipeline: representation enhancement, interaction graph enhancement, and optimization enhancement. The graph is built by fusing semantic and co-occurrence signals, then refined via Neighborhood Consistency Edge Reweighting (NCER), which applies the triadic closure principle to amplify structurally reliable edges and suppress spurious ones. Once constructed, the graph is leveraged in three complementary ways: (1) Item-item propagation with a Residual II Gate (RIG) that adaptively controls per-item absorption of semantic neighborhood signals for representation enhancement; (2) A content-guided UI graph expansion that introduces virtual user-item edges through high-confidence semantic neighbors for interaction graph enhancement; (3) II-Neighbor BPR Augmentation (INA) that treats top neighbors of positive items as discounted soft positives for optimization enhancement. We provide theoretical analysis showing that NCER reduces the spectral noise-to-signal ratio, RIG converges to a non-degenerate gating regime, and INA yields a tighter generalization bound. Extensive experiments on four datasets demonstrate that IIMRec consistently outperforms state-of-the-art baselines while running faster and consuming less GPU memory, with particularly strong gains under cold-start and sparse-interaction conditions.

Figures

Figures reproduced from arXiv: 2607.24607 by Edith C. H. Ngai, Jinfeng Xu, Jinze Li, Shujie Li, Shuo Yang, Yipeng Du, Zewei Liu, Zheyu Chen, Ziyue Peng.

Figure 1
Figure 1. Figure 1: The overall structure of IIMRec. capacity for cross-dimensional mapping. Since users lack multi￾modal raw features, user embeddings 𝑬 𝑚 𝑢 ∈ R |U |×𝑑 are randomly initialized for each modality 𝑚. Beyond the spatial (original) features, we extract spectral features via the Discrete Cosine Transform (DCT) to capture frequency￾domain patterns that are complementary to spatial representations: ˜𝒇𝑖 = Concat(DCT(… view at source ↗
Figure 2
Figure 2. Figure 2: Ablation study (Recall@20) on four datasets. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Sparsity analysis for IIMRec across all datasets. Tbl3Cldtt lill dtt [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Hyper-parameter analysis for IIMRec across all [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 4 linked inside Pith

  1. [1]

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

    2025

  2. [2]

    Zheyu Chen, Jinfeng Xu, Hewei Wang, Shuo Yang, Zitong Wan, and Haibo Hu

  3. [3]

    2010.Networks, crowds, and markets: Reasoning about a highly connected world

    David Easley, Jon Kleinberg, et al. 2010.Networks, crowds, and markets: Reasoning about a highly connected world. Vol. 1. Cambridge university press Cambridge

    2010

  4. [4]

    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

  5. [5]

    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

  6. [6]

    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

  7. [7]

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

  8. [8]

    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

  9. [9]

    Yijie Li, Hewei Wang, Jinfeng Xu, Puzhen Wu, Yunzhong Xiao, Shaofan Wang, and Soumyabrata Dev. 2025. Ddunet: Dual dynamic u-net for highly-efficient cloud segmentation. InIGARSS 2025-2025 IEEE International Geoscience and Remote Sensing Symposium. IEEE, 4705–4711

    2025

  10. [10]

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

  11. [11]

    Nagarajan Natarajan, Inderjit S Dhillon, Pradeep K Ravikumar, and Ambuj Tewari

  12. [12]

    Rongqing Kenneth Ong and Andy WH Khong. 2025. Spectrum-based modality representation fusion graph convolutional network for multimodal recommen- dation. InProceedings of the Eighteenth ACM International Conference on Web Search and Data Mining. 773–781

    2025

  13. [13]

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

    Pith/arXiv arXiv 2018

  14. [14]

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

  15. [15]

    Jinhui Tang, Xiaoyu Du, Xiangnan He, Fajie Yuan, Qi Tian, and Tat-Seng Chua

  16. [16]

    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

  17. [17]

    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

  18. [18]

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

    2023

  19. [19]

    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. [20]

    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. [21]

    Jinfeng Xu, Zheyu Chen, Jinze Li, Shuo Yang, Hewei Wang, Yijie Li, Mengran Li, Puzhen Wu, and Edith CH Ngai. 2025. Mdvt: Enhancing multimodal recommen- dation with model-agnostic multimodal-driven virtual triplets. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 3378–3389

    2025

  22. [22]

    Jinfeng Xu, Zheyu Chen, Jinze Li, Shuo Yang, Wei Wang, Xiping Hu, Raymond Chi-Wing Wong, and Edith CH Ngai. 2025. Enhancing Robustness and General- ization Capability for Multimodal Recommender Systems via Sharpness-Aware Minimization.IEEE Transactions on Knowledge and Data Engineering(2025)

    2025

  23. [23]

    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. InProceedings of the 48th In- ternational ACM SIGIR Conference on Research and Development in Information Retrieval. 1830–1839

    2025

  24. [24]

    Jinfeng Xu, Zheyu Chen, Shuo Yang, Jinze Li, and Edith CH Ngai. 2025. The Best is Yet to Come: Graph Convolution in the Testing Phase for Multimodal Recommendation. InProceedings of the 33rd ACM International Conference on Multimedia. 6325–6334

    2025

  25. [25]

    Jinfeng Xu, Zheyu Chen, Shuo Yang, Jinze Li, Zitong Wan, Hewei Wang, Weijie Liu, Yijie Li, and Edith CH Ngai. 2025. VI-MMRec: Similarity-Aware Training Cost-free Virtual User-Item Interactions for Multimodal Recommendation.arXiv preprint arXiv:2512.08702(2025)

    arXiv 2025

  26. [26]

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

  27. [27]

    Jinfeng Xu, Zheyu Chen, Shuo Yang, Jinze Li, Hewei Wang, Jianheng Tang, Wei Wang, Xiping Hu, and Edith CH Ngai. 2026. Well Begun is Half Done: Training-Free and Model-Agnostic Semantically Guaranteed User Representa- tion Initialization for Multimodal Recommendation. InProceedings of the 49th International ACM SIGIR Conference on Research and Development ...

    2026

  28. [28]

    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. InProceedings of the Nineteenth ACM Conference on Recommender Systems. 319–329

    2025

  29. [29]

    Jinfeng Xu, Zheyu Chen, Shuo Yang, Jinze Li, Hewei Wang, Wei Wang, Xip- ing Hu, and Edith Ngai. 2026. NLGCL+: Naturally Existing Neighbour Layers Graph Contrastive Learning with Adaptive Sample Weighting for Multimodal Recommendation.ACM Transactions on Recommender Systems(2026)

    2026

  30. [30]

    Jinfeng Xu, Zheyu Chen, Shuo Yang, Jinze Li, Wei Wang, Xiping Hu, Steven Hoi, and Edith Ngai. 2026. A survey on multimodal recommender systems: Recent advances and future directions.IEEE Transactions on Multimedia(2026)

    2026

  31. [31]

    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. InProceedings of the 45th international ACM SIGIR conference on research and development in information retrieval. 1294–1303

    2022

  32. [32]

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

  33. [33]

    InProceedings of the AAAI Conference on Artificial Intelligence, Vol

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

  34. [34]

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

    Pith/arXiv arXiv 2023

  35. [35]

    Xin Zhou, Donghui Lin, Yong Liu, and Chunyan Miao. 2023. Layer-refined graph convolutional networks for recommendation. In2023 IEEE 39th International Conference on Data Engineering (ICDE). IEEE, 1247–1259

    2023

  36. [36]

    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

  37. [37]

    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. MM ’26, November 10–14, 2026, Rio de Janeiro, Brazil Xu et al. A Proof of Theorem 5.2 (NCER Noise Reduction) Proof. Decompose the...

    2023

  38. [41]

    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)

    Pith/arXiv arXiv 2023

  39. [46]

    • SimGCL[ 31] performs graph contrastive learning by injecting random noise directly into the representation space instead of augmenting the graph structure

    General recommendation models: • LightGCN[ 6] simplifies graph convolutional network (GCN)- based collaborative filtering by eliminating redundant compo- nents such as feature transformation and nonlinear activation. • SimGCL[ 31] performs graph contrastive learning by injecting random noise directly into the representation space instead of augmenting the...

  40. [47]

    • DualGNN[ 17] constructs an auxiliary user-user graph to capture user preference correlations alongside multimodal item features

    Multimodal recommendation models: • MMGCN[ 20] employs separate GCNs to process and integrate information from different modalities. • DualGNN[ 17] constructs an auxiliary user-user graph to capture user preference correlations alongside multimodal item features. • LATTICE[ 32] builds item-item semantic graphs from multi- modal features to capture latent ...

    2026

  41. [2012]

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

    Pith/arXiv arXiv 2012

  42. [2013]

    Learning with noisy labels.Advances in neural information processing systems26 (2013)

    2013

  43. [2015]

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

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

  44. [2019]

    IEEE Transactions on Knowledge and Data Engineering32, 5 (2019), 855–867

    Adversarial training towards robust multimedia recommender system. IEEE Transactions on Knowledge and Data Engineering32, 5 (2019), 855–867

    2019

  45. [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

  46. [2024]

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

    2024

  47. [2025]

    InProceed- ings of the 34th ACM International Conference on Information and Knowledge Management

    Hypercomplex Prompt-aware Multimodal Recommendation. InProceed- ings of the 34th ACM International Conference on Information and Knowledge Management. 403–414