Pith. sign in

REVIEW 4 major objections 5 minor 60 references

CaIRec claims that incomplete multimodal recommendation requires both structurally calibrated imputation and ranking-oriented adaptation, reporting a 3.77% average gain over eleven baselines on three Amazon datasets.

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 ↗

CaIRec consistently beats eleven baselines on three Amazon datasets (about 2-7% relative gains) by combining latent imputation, spectral cross-modal calibration, and recommendation-space alignment of recovered item features.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection A clean, honest two-stage imputation method for missing modalities, but the headline gains are small and not verifiable without code or per-seed stats. the 4 major comments →

arxiv 2607.26720 v2 pith:CZB7LKJT submitted 2026-07-29 cs.IR

CaIRec: Calibrated Modality Imputation for Incomplete Multimodal Recommendation

Ruiyu Liu , Xiaohao Liu , Miaomiao Cai , Yunshan Ma , See-Kiong Ng This is my paper
classification cs.IR
keywords incomplete multimodal recommendationmodality imputationcross-modal structural distortionpreference adaptation gapshared latent modelitem graphpseudo-missing instancespersonalized ranking
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.

The reading

Multimodal recommenders use item content such as images and text, but in practice some items lack one of these modalities, and that missingness degrades ranking. CaIRec is a two-stage answer: Structural Imputation Calibration (SIC) estimates the missing modality from a shared latent factor inferred from the item's available modalities, then calibrates the within-item cross-modal organization; Preference-oriented Representation Calibration (PRC) then adapts the recovered representations and item neighborhoods to the ranking objective. The paper claims this pipeline consistently outperforms eleven baselines on three Amazon datasets under a fixed 50% missing mask, by an average relative 3.77% over the strongest baseline. Ablations attribute the gain to both stages, with removal of PRC causing the largest drop (up to 10.44% relative in Recall@20), and the method's advantage grows as the missing rate increases and concentrates on the incomplete items themselves.

Core claim

The central claim is that missing-modality representations are not enough on their own: they must be structurally calibrated within an item and then re-adapted to the recommendation space. CaIRec's SIC stage uses a conjugate Gaussian shared-latent model, where an item-level latent factor generates each modality through a linear map, so a missing modality is a conditional estimate from the posterior mean over observed modalities. Structural regularization on the Gram matrix of completed representations plus correspondence supervision from genuinely observed modality pairs corrects Cross-modal Structural Distortion. The PRC stage constructs pseudo-missing instances — temporarily masking an obs

What carries the argument

The load-bearing mechanism is the shared-latent imputation backbone followed by two calibration stages. The backbone is a probabilistic linear-Gaussian model: each item has a latent variable, each observed modality is generated as a Gaussian whose mean is a linear map of that latent variable, and missing modalities are estimated from the closed-form posterior mean. Around this, SIC adds structural calibration — spectral regularization of each item's modality Gram matrix, concentrating the leading eigenvalue and diversifying principal directions — and correspondence calibration, a same-item binary classification on genuinely observed cross-modal pairs. PRC then adds representation-level adapt

Load-bearing premise

The method rests on the assumption that a missing modality's projection is a linear function of a single item-level latent factor inferred from the observed modalities; if cross-modal predictability is weak — and the paper's own appendix shows image-to-text imputation is consistently worse, a gap the authors do not explain — then the imputed representations carry little real content and the subsequent calibration cannot recover it.

What would settle it

Measure the explained variance, or predictive R-squared, of the linear-Gaussian imputation on held-out observed modality pairs: if text features are largely unpredictable from image features in a given dataset, the backbone's premise fails and CaIRec's gains must come from elsewhere. A more direct experiment is to run CaIRec on a dataset engineered so that one modality shares no linear information with the other; under that condition, if the method still beats zero-fill baselines, the improvement is not attributable to imputed content.

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

If this is right

  • If correct, completion alone is insufficient; structural calibration and ranking-oriented adaptation are both necessary for incomplete multimodal recommendation.
  • Explicit completion methods generally outperform robustness-only methods, suggesting recommenders should estimate missing features rather than merely be invariant to their absence.
  • Gains concentrate on modality-incomplete items and widen as the missing rate grows, so the approach is most valuable in severe-missingness settings.
  • The SIC completion stage can be plugged into an existing recommender and improve it, so structurally calibrated imputation may transfer across backbones.
  • Stage-wise optimization, fixing SIC before PRC, beats joint training, implying the completion and ranking objectives are best separated.

Where Pith is reading between the lines

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

  • The paper's own results show text-imputed-from-image performs consistently worse than image-imputed-from-text; that asymmetry suggests the linear-Gaussian cross-modal predictor is the true bottleneck, and the unexplained gap likely reflects asymmetric information content between modalities rather than calibration quality.
  • Because the principal-direction term in the structural loss is inert in the two-modality setting, a real test of the structural calibration idea requires three or more modalities; until then, the empirical contribution of structural calibration rests mainly on the eigenvalue-concentration term.
  • The reliance on items with at least two observed modalities for correspondence and pseudo-missing objectives means the method will degrade in settings where most items are single-modality; a testable extension is to bootstrap supervision from clusters of similar items instead.
  • The stage-wise-versus-joint result hints at gradient interference between completion and ranking objectives; measuring gradient alignment during joint training would tell whether the difference is truly about interference or about optimization stability.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 5 minor

Summary. The paper proposes CaIRec, a two-stage framework for incomplete multimodal recommendation. The first stage (SIC) estimates missing-modality representations through a shared-latent linear-Gaussian imputation backbone, then calibrates within-item cross-modal structure via a spectral regularizer and a correspondence objective on observed modality pairs. The second stage (PRC) adapts the recovered representations to the ranking space through pseudo-missing alignment and constructs completion-aware item graphs that combine collaborative and completed-content similarities, followed by LightGCN-style propagation and BPR optimization. The paper reports consistent superiority over eleven baselines on three Amazon datasets under a fixed 50% modality-missing mask, with ablations attributing gains to both stages. Appendices provide closed-form posterior inference, complexity analysis, and an unusually candid discussion of limitations, including the bimodal degeneracy of the structural regularizer and direction-dependent imputation quality.

Significance. If the empirical claims hold, CaIRec would be a useful contribution to incomplete multimodal recommendation: it cleanly separates structurally calibrated imputation from recommendation-oriented adaptation, and the plug-in evaluation of SIC (Appendix C.4) suggests the completion component may transfer to other backbones. The paper is also transparent about its own limitations (Appendices A.4, A.7, C.1), which is commendable. However, the central claim is empirical, and the evidence as presented is not independently verifiable: Table 2 reports five-run means without standard deviations, per-seed values, or code release, and several margins over the strongest baseline are very small (e.g., Beauty R@10: 0.0529 vs. 0.0520, a 1.73% relative difference). The baseline tuning procedure is also not documented in a way that rules out under-tuned competitors. In addition, the paper's own Appendix A.4 shows that in the bimodal setting used in all experiments, the structural calibration term reduces to increasing the absolute cross-modal similarity |c_i|, which substantially weakens the claimed 'Cross-modal Structural Distortion' mechanism.

major comments (4)
  1. [§4.2, Table 2] The claim that CaIRec 'consistently achieves the strongest performance' is load-bearing but not supported by the reported statistics. Table 2 gives only five-run averages; no standard deviations, confidence intervals, per-seed values, or test statistics are provided. The text states that significance was assessed by a paired two-sided t-test at p<0.005, but no test statistic or degrees of freedom are reported. Several improvements over the strongest baseline are very small (e.g., Beauty R@10: 0.0529 vs. 0.0520; Clothing N@10: 0.0291 vs. 0.0286). With n=5 and full-ranking on 99.9%-sparse data, such margins could easily be within run-to-run noise. To support the central empirical claim, the authors should report per-seed results, standard deviations, and exact test statistics, and ideally release code and data splits so the comparison can be reproduced.
  2. [§4.1.3, Appendix B.2] The paper says 'all methods use the same split, mask, and hyperparameter search protocol,' but the protocol is not specified for baselines. Appendix B.2 lists only CaIRec's hyperparameter ranges and selection criteria; it does not state the search budgets, ranges, or final hyperparameters for LightGCN, MIG-GT, DGMRec, HEAT, or the other rivals. Without evidence that each baseline was tuned with comparable effort, the reported relative improvements could partly reflect under-tuned baselines. This is especially important for the strongest competitor MIG-GT, where the average margin is 3.77% but some cells are below 2%. The authors should provide per-baseline search ranges, selection criteria, and final configurations, or explain why equal tuning effort is guaranteed.
  3. [§3.1.2, Appendix A.4] The structural calibration mechanism is weaker than the paper's framing. Appendix A.4 admits that in the bimodal case the principal-direction term provides no additional continuous gradient within a fixed similarity-sign region, and the leading-eigenvalue term simply increases |c_i|, i.e., it encourages the two modality representations of each item to be more similar. All experiments in the paper are bimodal, so the 'cross-modal structural distortion' addressed by SIC is empirically a one-dimensional concentration effect, not the richer spectral organization suggested by Eq. (7). The authors should either present experiments with more than two modalities where the principal-direction term has nontrivial effect, or reinterpret the SIC gain as a similarity-based regularizer and compare directly against simpler alternatives (e.g., a cosine-similarity maximization or a margin loss between mo
  4. [§3.1.1 Eq. (3), Appendix C.1] The imputation backbone assumes a single linear-Gaussian mapping from a shared latent factor to each modality space. The paper's own single-modality missingness experiments show that text completion from images (I->T) is consistently much worse than image completion from text (T->I), and the authors state they 'cannot isolate' whether this reflects the utility of textual content or the difficulty of estimation. This is a mechanism-level concern: under the random missing mask used in the main experiments, roughly half of incomplete items lose text, so the aggregate gains may be driven by the easier direction. The authors should report full-model results (not only the w/o-SIC variant) separately by missing direction, and ideally analyze the quality of imputed representations (e.g., retrieval accuracy or downstream task performance) to clarify what content the recovered features actually ca
minor comments (5)
  1. [Eq. (23)] The denominator of the pseudo-recovered representation has a formatting artifact: '\r\r\rW^m b_i + mu^m\r\r\r' should be a clean norm notation. Please fix.
  2. [Figures 3, 4, and 10] The bar/line plots show only point estimates. Adding error bars (e.g., standard deviation over the five seeds) would help the reader assess whether the group-wise and missing-rate comparisons are reliable, especially where the plotted differences are small.
  3. [Appendix A.7, §4.1.2] Under high missing rates (e.g., 90%), only 10% of items retain both modalities, which limits the number of eligible pseudo-missing pairs for L_align and correspondence pairs for L_corr. The paper should quantify the number of eligible pairs at each missing rate, or discuss the effective supervision level. This would also clarify why SIC still helps in Figure 6.
  4. [§3.1.2] The term 'correspondence calibration' is used for an objective that resembles standard multimodal contrastive/alignment learning. Please clarify in the text how it differs from existing cross-modal alignment losses beyond the use of only observed pairs.
  5. [Appendix C.3] The stage-wise training comparison is useful, but the conclusion that fixing SIC 'provides a more stable input space' is speculative. The authors correctly note that gradient interference is not established as the only cause; consider adding a simple diagnostic (e.g., gradient norm or representation drift over training) to support the explanation.

Circularity Check

0 steps flagged

No significant circularity; the central claims are grounded in external benchmark comparisons, and the self-citations are not load-bearing.

full rationale

The paper's central claim is empirical: CaIRec outperforms eleven external baselines on three Amazon datasets under a fixed 50% modal-missing mask, with mean results over five runs reported in Table 2 and significance indicated by paired t-tests. The evaluation uses held-out user–item interactions, so no quantity used to fit parameters is also used as the evidence for the headline result. The pseudo-missing alignment (Lalign) is a training regularizer that maps recovered representations toward observed ones; it is not a fitted parameter later renamed as a prediction, and genuine missing-modality items are evaluated separately. The shared-latent imputation assumption (Eq. 3–6) is a modeling choice, not a circular reduction: it defines how missing representations are estimated, but the recommendation loss (Eq. 15) and downstream ranking metrics are independent of that definition. The paper does cite the authors' prior work—[27] for spectral structural regularization and [28] for calibrated recovery—and builds on [20, 28] in its framing, but these citations provide design components and motivation, not the empirical justification. The ablation study and comparisons against DGMRec, HEAT, and MIG-GT give external evidence for the method's effectiveness. Appendix A.6 explicitly distinguishes CaIRec from CalMRL, and Appendix A.7 / C.1 disclose limitations about generalization, the principal-direction component, and the unexplained I->T performance gap. These are robustness and correctness concerns, not circularity. The self-citations are therefore minor and not load-bearing for the central benchmark claim.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The paper introduces no new postulated entities: the pseudo-missing instances are a training construct (masked observed modalities), not an entity with independent falsifiable handles. The free parameters are the five hand-tuned weights (lambda_str, lambda_corr, lambda_align, lambda_reg, eta, k, tau are enumerable), the unspecified latent dimension d_beta, and the learned generative parameters of the imputation backbone. The axioms are the linear-Gaussian recoverability assumption, the O_i != empty scope assumption, the ad hoc concentration prior on cross-modal similarity, and the synthetic-missingness testbed assumption — the last three are acknowledged by the authors in A.4/A.7.

free parameters (5)
  • lambda_str, lambda_corr (SIC calibration weights) = lambda_str in {0.05,...,5.0}, lambda_corr in {0.01,...,0.5}, chosen by validation R@20
    Hand-tuned objective weights; the sensitivity analysis (Fig. 5) shows performance varies by up to ~3% relative across the grid, so the choice affects the reported gains.
  • lambda_align, lambda_reg (PRC weights) = lambda_align in {1e-4,...,0.1}, lambda_reg in {1e-3,...,0.02}, chosen by validation R@20
    Tuned on validation Recall@20 per Appendix B.2; the reported gains depend on these balances per Fig. 5.
  • eta (collaborative/content fusion), k (Top-k neighbors), tau (structural temperature) = eta in {0.1,...,0.9}, k in {1,...,20}, tau in {0.05,...,0.5}
    Graph and temperature hyperparameters selected on the validation split (Appendix B.2, Fig. 9); the shown increase-then-decrease curves indicate sensitivity to k and eta.
  • d_beta (shared-latent dimension) = unspecified in the text
    The latent dimension of beta_i (Eq. 3) is never listed in Appendix B.2, though it directly controls the capacity of the imputation backbone.
  • Generative parameters W^m, mu^m, sigma^m = learned via Eq. (5) on observed modalities
    These define the imputation mapping (Eq. 6); they are fitted model parameters, standard for the method, but they are the quantities that determine what information the imputed features carry.
axioms (4)
  • domain assumption Missing modalities are recoverable from a linear-Gaussian shared latent factor (Eq. 3: p(z^m|beta)=N(W^m beta+mu^m,(sigma^m)^2 I)) with per-modality projections (Eq. 2).
    Load-bearing for SIC: if the shared latent carries little predictable cross-modal information, the recovered representations approach modality-mean vectors and downstream calibration cannot add content information. The paper's own C.1 results show I->T (text imputed from image) performs worse, with the cause left unidentified.
  • domain assumption Every item has at least one observed modality (O_i != empty).
    Stated in Preliminaries (Section 2). Items missing all modalities cannot be imputed, and the evaluation never tests this case, so the method's scope excludes it.
  • ad hoc to paper Concentrating the within-item cross-modal relation (increasing |c_i| in the bimodal Gram matrix) improves recommendation (Eq. 7, Appendix A.4).
    The leading-eigenvalue term is a hand-chosen prior; the paper itself states (A.4) it does not determine the sign or semantic correctness of the pairwise relation and acts only as a relation-concentration regularizer. Its support is the w/o StrCal ablation (-0.85% to -5.16%).
  • domain assumption Synthetic random single-modality missingness (50% of items, one modality removed uniformly) is a valid testbed for the robustness claims.
    All experiments use simulated masks fixed across train/validation/test. Real missingness may depend on popularity, category, or content quality; the authors acknowledge this in A.7, limiting the scope of the claims.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CaIRec: Calibrated Modality Imputation for Incomplete Multimodal Recommendation." pith.science (2026). https://pith.science/paper/CZB7LKJT

@misc{pith2026260726720,
  author       = {Pith},
  title        = {Pith review of: CaIRec: Calibrated Modality Imputation for Incomplete Multimodal Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CZB7LKJT}},
  note         = {Machine review of arXiv:2607.26720}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Real-world multimodal recommender systems often face incomplete modality observations, where items lack images, text, or other content features. Such incompleteness weakens item representations and degrades recommendation performance. Existing modality imputation methods estimate missing representations from available item content, but two challenges remain. First, they optimize the recovered representation itself without explicitly considering its relations with other modalities of the same item. The completed modalities may therefore form inconsistent cross-modal relations, causing Cross-modal Structural Distortion. Second, even structurally coherent recovered information may remain ineffective for personalized ranking. Recovered representations receive limited ranking-oriented guidance, while modality missingness disrupts the item neighborhoods required for preference propagation, resulting in a Preference Adaptation Gap. To address these challenges, we propose Calibrated Imputation for Incomplete Multimodal Recommendation (CaIRec), a two-stage framework. Structural Imputation Calibration (SIC) estimates missing-modality representations from shared information inferred from available modalities and calibrates their cross-modal organization through structural regularization and correspondence supervision from observed modality pairs. Preference-oriented Representation Calibration (PRC) performs recommendation-specific adaptation at both the representation and relation levels. It constructs pseudo-missing instances to align recovered representations with observed counterparts shaped by ranking supervision in the recommendation space. It further builds completion-aware item graphs by integrating completed content relations with collaborative evidence. Extensive experiments on three datasets under different modality-missing settings demonstrate the effectiveness and robustness of CaIRec.

Figures

Figures reproduced from arXiv: 2607.26720 by Miaomiao Cai, Ruiyu Liu, See-kiong Ng, Xiaohao Liu, Yunshan Ma.

Figure 1
Figure 1. Figure 1: The two challenges in incomplete multimodal rec [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed CaIRec framework. SIC estimates missing-modality representations and calibrates their within-item cross-modal organization. PRC then adapts recovered representations in the recommendation space and supple￾ments item relations for preference propagation and personalized ranking. second term regularizes structural variation across items, while the squared inner product removes the si… view at source ↗
Figure 3
Figure 3. Figure 3: Performance across item groups. removing the completion-aware item graph reduces connections available for preference propagation. These results support the representation- and relation-level design of PRC. The ablation re￾sults validate the sequential design of CaIRec, where SIC calibrates the within-item cross-modal organization and PRC adapts repre￾sentations and item neighborhoods to personalized ranki… view at source ↗
Figure 5
Figure 5. Figure 5: examines the objective weights in SIC and PRC. For SIC, moderate values of 𝜆str and 𝜆corr generally yield stronger perfor￾mance. A weak structural constraint provides limited cross-modal regularization, whereas an excessively large value may over-constrain the completion space. The effect of correspondence calibration also depends on structural calibration, suggesting that relation-pattern regularization a… view at source ↗
Figure 6
Figure 6. Figure 6: Performance of CaIRec and its w/o SIC variant under varying single-modality missing rates. 𝑇→𝐼 denotes image completion from text, while 𝐼→𝑇 denotes text completion from images. 10% 30% 50% 70% 90% Missing Rate 0.0320 0.0340 0.0360 0.0380 0.0400 NDCG@20 (a) Clothing 10% 30% 50% 70% 90% Missing Rate 0.0700 0.0750 0.0800 0.0850 0.0900 Recall@20 (b) Clothing 10% 30% 50% 70% 90% Missing Rate 0.0420 0.0440 0.04… view at source ↗
Figure 7
Figure 7. Figure 7: Additional recommendation results under varying modality-missing rates. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Performance of PGL and PGL+SIC on the three [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Additional hyperparameter sensitivity on Clothing. [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗

Review history (2 revisions) →

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

60 extracted references · 7 linked inside Pith

  1. [1]

    Bach and Michael I

    Francis R. Bach and Michael I. Jordan. 2005. A Probabilistic Interpretation of Canonical Correlation Analysis.Technical Report(2005)

    2005

  2. [2]

    Haoyue Bai, Le Wu, Min Hou, Miaomiao Cai, Zhuangzhuang He, Yuyang Zhou, Richang Hong, and Meng Wang. 2024. Multimodality Invariant Learning for Multimedia-Based New Item Recommendation.SIGIR(2024), 677–686

    2024

  3. [3]

    Huilin Chen, Miaomiao Cai, Fan Liu, Zhiyong Cheng, Richang Hong, and Meng Wang. 2025. I 3-MRec: Invariant Learning with Information Bottleneck for In- complete Modality Recommendation.ACM MM(2025), 6133–6142

    2025

  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.SIGIR(2017), 335–344

    2017

  5. [5]

    Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. 2023. VAST: A Vision-Audio-Subtitle-Text Omni-Modality Founda- tion Model and Dataset.NeurIPS(2023), 72842–72866

    2023

  6. [6]

    Giordano Cicchetti, Eleonora Grassucci, and Danilo Comminiello. 2025. A TRIAN- GLE Enables Multimodal Alignment Beyond Cosine Similarity.NeurIPS(2025)

    2025

  7. [7]

    Christian Ganhör, Marta Moscati, Anna Hausberger, Shah Nawaz, and Markus Schedl. 2024. A Multimodal Single-Branch Embedding Network for Recommen- dation in Cold-Start and Missing Modality Scenarios.arXiv:2409.17864(2024)

    Pith/arXiv arXiv 2024

  8. [8]

    Benyamin Ghojogh, Ali Ghodsi, Fakhri Karray, and Mark Crowley. 2021. Factor Analysis, Probabilistic Principal Component Analysis, Variational Inference, and Variational Autoencoder: Tutorial and Survey.arXiv:2101.00734(2021)

    Pith/arXiv arXiv 2021

  9. [9]

    Zhiqiang Guo, Jianjun Li, Guohui Li, Chaoyang Wang, Si Shi, and Bin Ruan. 2024. LGMRec: Local and Global Graph Learning for Multimodal Recommendation. AAAI(2024), 8454–8462

    2024

  10. [10]

    Ruining He and Julian McAuley. 2016. VBPR: Visual Bayesian Personalized Ranking from Implicit Feedback.AAAI(2016), 144–150

    2016

  11. [11]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation.SIGIR(2020), 639–648

    2020

  12. [12]

    Jun Hu, Bryan Hooi, Bingsheng He, and Yinwei Wei. 2025. Modality-Independent Graph Neural Networks with Global Transformers for Multimodal Recommen- dation.AAAI(2025), 11790–11798

    2025

  13. [13]

    Yifan Hu, Yehuda Koren, and Chris Volinsky. 2008. Collaborative Filtering for Implicit Feedback Datasets.ICDM(2008), 263–272

    2008

  14. [14]

    Kalervo Järvelin and Jaana Kekäläinen. 2002. Cumulated Gain-Based Evaluation of IR Techniques.TOIS(2002), 422–446

    2002

  15. [15]

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

  16. [16]

    Wang-Cheng Kang, Chen Fang, Zhaowen Wang, and Julian McAuley. 2017. Visually-Aware Fashion Recommendation and Design with Generative Image Models.ICDM(2017), 207–216

    2017

  17. [17]

    Jiwan Kim, Hongseok Kang, Sein Kim, Kibum Kim, and Chanyoung Park. 2025. Disentangling and Generating Modalities for Recommendation in Missing Modal- ity Scenarios.SIGIR(2025), 1820–1829

    2025

  18. [18]

    Arto Klami, Seppo Virtanen, Eemeli Leppäaho, and Samuel Kaski. 2015. Group Factor Analysis.TNNLS(2015), 2136–2147

    2015

  19. [19]

    Jin Li, Shoujin Wang, Qi Zhang, Shui Yu, and Fang Chen. 2025. Generating with Fairness: A Modality-Diffused Counterfactual Framework for Incomplete Multimodal Recommendations.WWW(2025), 2787–2798

    2025

  20. [20]

    Yuan Li, Jun Hu, Jiaxin Jiang, Bryan Hooi, and Bingsheng He. 2026. Robust Mul- timodal Recommendation via Graph Retrieval-Enhanced Modality Completion. arXiv:2605.00670(2026)

    Pith/arXiv arXiv 2026

  21. [21]

    Kankanhalli

    Zhenyang Li, Fan Liu, Yinwei Wei, Zhiyong Cheng, Liqiang Nie, and Mohan S. Kankanhalli. 2024. Attribute-Driven Disentangled Representation Learning for Multimodal Recommendation.ACM MM(2024), 9660–9669

    2024

  22. [22]

    Guojiao Lin, Zhen Meng, Dongjie Wang, Qingqing Long, Yuanchun Zhou, and Meng Xiao. 2024. GUME: Graphs and User Modalities Enhancement for Long-Tail Multimodal Recommendation.arXiv:2407.12338(2024)

    Pith/arXiv arXiv 2024

  23. [23]

    Xixun Lin, Rui Liu, Yanan Cao, Lixin Zou, Qian Li, Yongxuan Wu, Yang Liu, Dawei Yin, and Guandong Xu. 2025. Contrastive Modality-Disentangled Learning for Multimodal Recommendation.TOIS(2025), 1–31

    2025

  24. [24]

    Zhenghong Lin, Yanchao Tan, Yunfei Zhan, Weiming Liu, Fan Wang, Chaochao Chen, Shiping Wang, and Carl Yang. 2023. Contrastive Intra- and Inter-Modality Generation for Enhancing Incomplete Multimedia Recommendation.ACM MM (2023), 6234–6242

    2023

  25. [25]

    Kankanhalli

    Fan Liu, Huilin Chen, Zhiyong Cheng, Anan Liu, Liqiang Nie, and Mohan S. Kankanhalli. 2023. Disentangled Multimodal Representation Learning for Rec- ommendation.TMM(2023), 7149–7159

    2023

  26. [26]

    Kankanhalli

    Fan Liu, Huilin Chen, Zhiyong Cheng, Liqiang Nie, and Mohan S. Kankanhalli

  27. [27]

    Xiaohao Liu, Xiaobo Xia, See-Kiong Ng, and Tat-Seng Chua. 2026. Principled Multimodal Representation Learning.TPAMI(2026), 9114–9128. Conference’17, July 2017, Washington, DC, USA Ruiyu Liu, Xiaohao Liu, Miaomiao Cai, Yunshan Ma, and See-Kiong Ng

    2026

  28. [28]

    Xiaohao Liu, Xiaobo Xia, Jiaheng Wei, Shuo Yang, Xiu Su, See-Kiong Ng, and Tat- Seng Chua. 2026. Calibrated Multimodal Representation Learning with Missing Modalities.ICML(2026)

    2026

  29. [29]

    Malliaros

    Daniele Malitesta, Emanuele Rossi, Claudio Pomo, Tommaso Di Noia, and Fragkiskos D. Malliaros. 2024. Do We Really Need to Drop Items with Miss- ing Modalities in Multimodal Recommendation?CIKM(2024), 3943–3948

    2024

  30. [30]

    Malliaros

    Daniele Malitesta, Emanuele Rossi, Claudio Pomo, Tommaso Di Noia, and Fragkiskos D. Malliaros. 2026. Training-Free Graph-Based Imputation of Missing Modalities in Multimodal Recommendation.TKDE(2026), 3250–3263

    2026

  31. [31]

    Malliaros, and Tommaso Di Noia

    Daniele Malitesta, Emanuele Rossi, Claudio Pomo, Fragkiskos D. Malliaros, and Tommaso Di Noia. 2024. Dealing with Missing Modalities in Multimodal Recom- mendation: A Feature Propagation-Based Approach.arXiv:2403.19841(2024)

    Pith/arXiv arXiv 2024

  32. [32]

    Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton van den Hengel

  33. [33]

    Jiquan Ngiam, Aditya Khosla, Mingyu Kim, Juhan Nam, Honglak Lee, and An- drew Y. Ng. 2011. Multimodal Deep Learning. InICML. 689–696

    2011

  34. [34]

    Yongxin Ni, Yu Cheng, Xiangyan Liu, Junchen Fu, Youhua Li, Xiangnan He, Yongfeng Zhang, and Fajie Yuan. 2023. A Content-Driven Micro-Video Recom- mendation Dataset at Scale.arXiv:2309.15379(2023)

    Pith/arXiv arXiv 2023

  35. [35]

    Rongqing Kenneth Ong and Andy W. H. Khong. 2025. Spectrum-Based Modality Representation Fusion Graph Convolutional Network for Multimodal Recom- mendation.WSDM(2025), 773–781

    2025

  36. [36]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings Using Siamese BERT-Networks.EMNLP-IJCNLP(2019), 3982–3992

    2019

  37. [37]

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

  38. [38]

    Nitish Srivastava and Ruslan Salakhutdinov. 2012. Multimodal Learning with Deep Boltzmann Machines. InNeurIPS. 2222–2230

    2012

  39. [39]

    Zhulin Tao, Xiaohao Liu, Yewei Xia, Xiang Wang, Lifang Yang, Xianglin Huang, and Tat-Seng Chua. 2023. Self-Supervised Learning for Multimedia Recommen- dation.TMM(2023), 5107–5116

    2023

  40. [40]

    Zhi Tao, Yinwei Wei, Xiang Wang, Xiangnan He, Xianglin Huang, and Tat-Seng Chua. 2020. MGAT: Multimodal Graph Attention Network for Recommendation. IPM(2020), 102277

    2020

  41. [41]

    Tipping and Christopher M

    Michael E. Tipping and Christopher M. Bishop. 1999. Probabilistic Principal Component Analysis.JRSS-B(1999), 611–622

    1999

  42. [42]

    Cheng Wang, Mathias Niepert, and Hui Li. 2018. LRMM: Learning to Recommend with Missing Modalities.EMNLP(2018), 3360–3370

    2018

  43. [43]

    Qifan Wang, Yinwei Wei, Jianhua Yin, Jianlong Wu, Xuemeng Song, and Liqiang Nie. 2023. DualGNN: Dual Graph Neural Network for Multimedia Recommenda- tion.TMM(2023), 1074–1084

    2023

  44. [44]

    Yinwei Wei, Wenqi Liu, Fan Liu, Xiang Wang, Liqiang Nie, and Tat-Seng Chua

  45. [45]

    Yinwei Wei, Xiang Wang, Liqiang Nie, Xiangnan He, and Tat-Seng Chua. 2020. GRCN: Graph-Refined Convolutional Network for Multimedia Recommendation with Implicit Feedback.ACM MM(2020), 3541–3549

    2020

  46. [46]

    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.ACM MM(2019), 1437–1445

    2019

  47. [47]

    Jinfeng Xu, Zheyu Chen, Shuo Yang, Jinze Li, Hewei Wang, and Edith C.-H. Ngai. 2024. MENTOR: Multi-Level Self-Supervised Learning for Multimodal Recommendation.arXiv:2402.19407(2024)

    Pith/arXiv arXiv 2024

  48. [48]

    Zixuan Yi, Xi Wang, Iadh Ounis, and Craig Macdonald. 2022. Multi-modal Graph Contrastive Learning for Micro-video Recommendation.SIGIR(2022), 1807–1811

    2022

  49. [49]

    SIGIR(2023), 1508–1517

    LightGT: A Light Graph Transformer for Multimedia Recommendation. SIGIR(2023), 1508–1517

    2023

  50. [50]

    Penghang Yu, Zhiyi Tan, Guanming Lu, and Bing-Kun Bao. 2023. Multi-View Graph Convolutional Network for Multimedia Recommendation.ACM MM (2023), 6576–6585

    2023

  51. [51]

    Penghang Yu, Zhiyi Tan, Guanming Lu, and Bing-Kun Bao. 2025. Mind Individual Information! Principal Graph Learning for Multimedia Recommendation.AAAI (2025), 13096–13105

    2025

  52. [52]

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

  53. [53]

    Xin Zhou and Zhiqi Shen. 2023. A Tale of Two Graphs: Freezing and Denoising Graph Structures for Multimodal Recommendation.ACM MM(2023), 935–943

    2023

  54. [54]

    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.SIGIR(2022), 1294–1303

    2022

  55. [60]

    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.WWW(2023), 845–854. A Additional Method Details The shared-latent imputation backbone in SIC follows conjugate Gaussian latent-variable modeling [1, 8, 18, 41]. This section pro- vides...

    2023

  56. [2009]

    BPR: Bayesian Personalized Ranking from Implicit Feedback.UAI(2009), 452–461

    2009

  57. [2015]

    Image-Based Recommendations on Styles and Substitutes.SIGIR(2015), 43–52

    2015

  58. [2021]

    Mining Latent Structures for Multimedia Recommendation.ACM MM (2021), 3872–3880

    2021

  59. [2023]

    ACM MM(2023), 6567–6575

    Semantic-Guided Feature Distillation for Multimodal Recommendation. ACM MM(2023), 6567–6575

    2023

  60. [2024]

    DiffMM: Multi-Modal Diffusion Model for Recommendation.ACM MM (2024), 7591–7599

    2024

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.