Pith. sign in

REVIEW 5 major objections 5 minor 4 cited by

RecGOAT claims that aligning LLM-derived modality features with ID-based collaborative signals through instance-level contrastive learning and distribution-level optimal transport yields a fused representation whose recommendation error is

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 · deepseek-v4-flash

2026-08-03 05:55 UTC pith:B4JTIBYH

load-bearing objection RecGOAT's empirical recipe looks plausible, but its headline theory doesn't connect to the implemented OT — Eq. (10)-(11) is a feature-wise transport, not the item-level Wasserstein distance used in Theorem 3.5, and the abstract overclaims 'strictly lower target error'. the 5 major comments →

arxiv 2602.00682 v2 pith:B4JTIBYH submitted 2026-01-31 cs.IR cs.AI

RecGOAT: Graph Optimal Adaptive Transport for LLM-Enhanced Multimodal Recommendation with Dual Semantic Alignment

Yuecheng Li , Hengwei Ju , Zeyu Song , Wei Yang , Chi Lu , Peng Jiang , Kun Gai This is my paper
classification cs.IR cs.AI
keywords multimodal recommendationlarge language modelssemantic alignmentoptimal transportWasserstein distancecontrastive learninggraph neural networksrepresentation fusion
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.

RecGOAT tries to establish that the reason large-language-model features hurt multimodal recommendation is a distribution mismatch: generative model embeddings and ID-based collaborative embeddings live in different semantic spaces, so naive fusion degrades performance. The paper's solution is a dual alignment: cross-modal contrastive learning at the instance level and an optimal adaptive transport step at the distribution level that moves each modality's feature distribution into the ID embedding space. Its central theoretical result, Theorem 3.5, bounds the fused representation's target error by the minimum over modalities of the single-modality error plus a term involving the 1-Wasserstein distance and the InfoNCE loss. The authors argue this proves both alignment consistency and fusion comprehensiveness, and report state-of-the-art results on three Amazon datasets. A sympathetic reader would care because it offers a principled, provably-grounded way to inject LLM knowledge into recommender systems without performance collapse.

Core claim

The paper's central claim is that distribution-level semantic alignment is the missing ingredient in LLM-enhanced multimodal recommendation. It constructs attentive graphs over item-item, user-item, and user-user relations using LLM/LVM embeddings, initializes user representations by prompting a large language model to summarize interaction histories, and then aligns the resulting modality features to ID embeddings at two granularities. At the instance level, a cross-modal InfoNCE loss pulls together representations of the same item across text, vision, and ID; at the distribution level, an optimal adaptive transport plan moves the modality feature distribution toward the ID distribution, wi

What carries the argument

The load-bearing object is the optimal adaptive transport (OAT) plan T^m = T^m_0 + tilde T^m, where T^m_0 is the entropy-regularized Sinkhorn solution to the discrete transport problem between modality m's feature distribution and the ID embedding distribution, and tilde T^m is a learned residual matrix that lets the purely geometric plan adapt to the downstream ranking task. The transported features are computed as Z_hat^m = Z^m T^m and fused with ID embeddings by a weighted average, producing the unified item representation. The theoretical argument rests on the 1-Wasserstein distance as a distribution-level cost and the InfoNCE contrastive loss as an instance-level cost; Lemmas 3.3 and 3.

Load-bearing premise

The proof assumes the optimal-transport objective being minimized is the 1-Wasserstein distance between item-level distributions, but the implemented cost matrix compares feature-dimension coordinates rather than item samples, so the trained model may not actually minimize the quantity that the theorem bounds.

What would settle it

On a held-out set, compute the empirical 1-Wasserstein distance between the modality-specific item distributions and the ID distribution as actually implemented (using the d x d cost matrix), and compare the unified representation's recommendation error against the right-hand side of Theorem 3.5; if the error exceeds the bound for any user embedding, the theorem's premise is violated. Alternatively, replace the transport cost with a true B x B item-level cost and check whether the empirical bound tightens, isolating whether the mismatch matters.

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

If this is right

  • If the theorem holds, practitioners can safely add LLM-derived features to a recommender by aligning distributions rather than by naive concatenation, which the paper shows can degrade performance.
  • The bound makes the alignment terms directly actionable: reducing either the Wasserstein distance or the InfoNCE loss should shrink the gap between the fused representation's error and the best single modality's error.
  • Because the guarantee holds for every modality and takes the minimum, adding a new modality (e.g., audio or video) cannot worsen the error bound below the best single modality, provided the alignment terms are controlled.
  • The framework implies that distribution-level alignment is essential, not optional: instance-level contrastive learning alone leaves a Wasserstein term uncontrolled, which the ablations corroborate by showing OAT outperforms CMCL alone.

Where Pith is reading between the lines

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

  • The abstract's 'strictly lower' wording goes beyond Theorem 3.5, which establishes an upper bound with an additive alignment penalty; the theorem as stated permits the fused error to equal the best modality's error plus a positive term.
  • The implementation's cost matrix in Eqs. (10)-(11) is d x d over feature dimensions, whereas the theorem's Wasserstein distance is between item-level distributions; if these are indeed different objects, the trained objective may not be the quantity the proof bounds.
  • The manuscript defers the online advertising deployment section until after acceptance, so the abstract's industrial-scalability claim is not evidenced in this version.
  • A natural testable extension is to add a third modality and verify empirically that the fused error remains bounded by the minimum over all single modalities plus the alignment term, as the theorem predicts.

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 / 5 minor

Summary. RecGOAT proposes a multimodal recommendation model that injects LLM/LVM embeddings into graph-based collaborative filtering, then applies a dual alignment: instance-level cross-modal contrastive learning (CMCL) and distribution-level optimal adaptive transport (OAT) with learnable residual matrices. The paper claims a theorem (Theorem 3.5) guaranteeing that the fused representation's target error is bounded by the minimum over modalities of the single-modality error plus Wasserstein and contrastive-loss terms, and it reports state-of-the-art results on the Baby, Sports, and Electronics Amazon datasets. The online-deployment experiment advertised in the abstract and introduction is, however, absent from the manuscript body.

Significance. If the theoretical results were valid, the paper would contribute a principled way to combine LLM semantic features with ID-based collaborative signals, with explicit finite-sample-style bounds. The empirical recipe appears plausible, and the release of code is a strength. However, the mismatch between the implemented OT objective and the theoretical Wasserstein distance, the unproved key lemma, the quantifier error in the main theorem, and the actual absence of the online experiment together undermine the paper's central claims. The advertised 'rigorous guarantees' do not follow from the presented analysis, and the claimed industrial validation is not part of the submission.

major comments (5)
  1. [§3.2.2, Eqs. (10)-(11) vs. §3.3.2, Theorem 3.5] The OT objective minimized in Eqs. (10)-(11) is not the 1-Wasserstein distance W1(P_m,Q_id) used in Theorem 3.5. C^m is defined as a d×d matrix whose (i,j) entry averages |Z^m_{b,i} − Z^id_{b,j}| over the batch, and the transport plan T is d×d with uniform marginals p,q. For W1 between the empirical item distributions P_m = (1/B)∑δ_{z^m_i} and Q_id = (1/B)∑δ_{z^id_i}, the coupling must be a B×B matrix with row/column sums 1/B, and the cost must be a B×B item-pair matrix. A d×d coupling on feature coordinates is not a joint distribution over items. Therefore the quantity optimized by OAT is not the W1 appearing in Eqs. (16)-(18), and Theorem 3.5 does not apply to the trained model. This is a load-bearing identification error, not a notational slip.
  2. [§3.3.2, Lemma 3.3] Lemma 3.3 is asserted without proof, yet it is the only bridge from the contrastive loss to the Euclidean distance between modality-specific and unified representations. The claimed bound E_i‖z^m_i−z_i‖ ≤ √(2τ L_CMCL + 2τ log B) is not a standard consequence of InfoNCE: cosine-similarity contrastive losses bound angles/cosine distances, not Euclidean distances, unless all representations are normalized and the unified representation is among the contrastive anchors; moreover the term 2τ log B is dimensionally and structurally unmotivated. Since Theorem 3.5 relies directly on this lemma, the proof is incomplete.
  3. [§3.3.2, Theorem 3.5 proof, Eq. (17)] The proof of the Consistency Guarantee has a quantifier error. After showing for each m that ε_m(f)−ε_F(f) ≤ (K+L*)W1(P_m,Q_id)+..., the text writes max_m ε_m(f)−ε_F(f) ≤ (K+L*)W1(P_m,Q_id)+..., leaving m free on the right. A valid uniform bound requires max_m of the right-hand side. As written Eq. (17) is ill-formed. The Comprehensiveness Guarantee Eq. (18) has a correct min over m, but given the OAT mismatch it still refers to a quantity the algorithm does not optimize.
  4. [Abstract and §4.5] The abstract and introduction claim that deployment on a large-scale online advertising platform validates industrial scalability, but §4.5 contains only the sentence 'We will present this section after the paper is accepted (subject to company review).' No online experiment, A/B test, or deployment result is included. This is a claimed contribution with no supporting evidence in the manuscript.
  5. [Abstract and Theorem 3.5] The abstract states that the unified representation 'achieves strictly lower target error than any single-modality representation,' but Theorem 3.5(2) only gives ε_F(f) ≤ min_m {ε_m(f) + (K+L*)W1(P_m,Q_id) + √(2τ(K+L*)^2(L_CMCL+log B))}. Since the added terms are positive in general, this does not imply strict dominance; the theorem is a triangle-inequality-style upper bound. Moreover, because the bound is expressed in terms of precisely the two quantities the training objective minimizes, the guarantee is close to a restatement of the objective rather than an externally grounded prediction. The abstract's claim is stronger than what is proven.
minor comments (5)
  1. [§3.2.2, Eq. (13)] The notation 'ˆZ^m = Z^m · T^m' describes a linear transformation on feature coordinates; calling it 'transporting each feature toward the ID embedding space' is misleading, since T^m is not a coupling of item distributions. Please clarify the geometric interpretation.
  2. [§3.3.1, Theorem 3.5 notation] M is defined as {t,v,id}, but for m=id, P_m=Q_id, making W1(P_id,Q_id)=0. The interpretation of 'modality-specific error' for m=id should be made explicit, especially in Eq. (18).
  3. [Figure 1] The caption reports performance improvements of 59% and 70% without giving absolute metric values or error bars. The figure is hard to interpret as a quantitative comparison.
  4. [§4.1.2 and Table 1] The text says metrics are averaged over 10 runs and reports a t-test, but Table 1 shows no standard deviations or confidence intervals. Please include variance information or explain why it is omitted.
  5. [General] The header uses '2018' in the ACM copyright and 'Under Review'; these are template artifacts. Also, several reference entries (e.g., [3], [7]) have formatting issues. The manuscript should be cleaned before resubmission.

Circularity Check

0 steps flagged

No significant circularity: the main theorem is a conditional Lipschitz/triangle-inequality bound, not a fitted prediction; the serious problems are proof gaps and a Wasserstein-objective mismatch, which are correctness issues rather than circularity.

full rationale

I walked the claimed derivation chain. Theorem 3.5 is a conditional bound: if the optimized transport cost W1(P_m,Q_id) and the InfoNCE loss L_CMCL are small, then the fused error is close to the best single-modality error. This follows from Lipschitz/triangle-inequality arguments and a standard InfoNCE distance bound; it does not fit a parameter to the quantity it then 'predicts', nor does it assume its own conclusion. The bound is weak and largely restates the training objective, but under the definitions of circularity used here (Eq. X = Eq. Y by construction, fitted parameter renamed as prediction, load-bearing self-citation), that is not a circular step. I found no load-bearing self-citation: [44]-[46] appear in related work and do not supply a uniqueness theorem or ansatz on which the main claim depends. The serious defects are correctness/validity issues, not circularity: (i) Eq. (10) defines C^m in R^{d x d}, while Eq. (11) claims this is W1(P_m,Q_id) for empirical item distributions, which requires a B x B coupling; the theorem's W1 is therefore not the quantity actually minimized; (ii) Lemmas 3.3 and 3.4 are asserted without proof; (iii) the proof of Theorem 3.5 leaves m free on the right-hand side after taking max_m on the left; (iv) the abstract's 'strictly lower target error' is stronger than Eq. (18), which gives only an upper bound with a nonnegative alignment term; (v) Section 4.5 states 'We will present this section after the paper is accepted', so the advertised online deployment evidence is absent. These concerns matter, but none makes the derivation equivalent to its inputs. Score 1 reflects the minor presence of non-load-bearing self-citation and the near-tautological character of the bound; the paper is not circular in the sense of the seven enumerated patterns.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 1 invented entities

The central theoretical claim rests on two unproved or misidentified pieces: Lemma 3.3's InfoNCE-distance bound and the identification of Eq (11)'s d x d OT with the item-level W1 in Theorem 3.5. Smoothness assumptions on user embeddings and the true preference function are stated but not validated. The learned residual transport matrix and fusion weights are fitted to data.

free parameters (6)
  • InfoNCE temperature tau = not reported
    Hyperparameter in Eq (7); it enters Lemma 3.3 and Theorem 3.5 and controls the claimed bound.
  • OT cost scaling factor s = not reported
    Introduced in Eq (10) to scale the cost matrix; affects the W1 values and Sinkhorn plan.
  • Fusion weights gamma_t, gamma_v = tuned on validation (Fig 3a)
    Weights in Eq (14); the paper shows a heatmap over them, so they are fitted per dataset.
  • KNN graph top-K = not reported
    Number of neighbors kept in item-item and user-user graphs (Eq 2), chosen by hand.
  • GAT heads H, LightGCN layers L, embedding dim d = not reported
    Architecture hyperparameters that affect all representations and the OT cost matrix size.
  • Residual transport matrices tilde-T^m = learned during training
    The adaptive part of the transport plan in Eq (12); fitted on the recommendation objective and not part of the proved W1 guarantee.
axioms (6)
  • domain assumption Assumption 3.1: user embeddings are fixed and bounded, ||u|| <= K
    Used to make the score function K-Lipschitz in the item representation; reasonable but unverified.
  • domain assumption Assumption 3.2: true preference function f* is L*-Lipschitz with respect to item representation
    Smoothness of preferences is assumed without measurement or evidence.
  • ad hoc to paper Lemma 3.3: E_i ||z^m_i - z_i|| <= sqrt(2 tau L_CMCL + 2 tau log B)
    Asserted without proof; not a standard direct consequence of InfoNCE and appears to ignore negative-similarity configurations.
  • ad hoc to paper Eq (10)-(11) compute the 1-Wasserstein distance between item-level distributions P_m and Q_id
    This identification is false as written: the cost matrix is d x d over feature dimensions, not B x B over item samples.
  • domain assumption Optimizing BPR, CMCL, and OAT reduces the true expected error epsilon_F
    The theorem bounds epsilon_F in terms of epsilon_m plus alignment losses, but the model is trained with BPR and no generalization argument is supplied.
  • standard math Sinkhorn-Knopp converges to the entropy-regularized optimal transport plan
    Standard result [9,32], used to compute T_0^m.
invented entities (1)
  • Task-adaptive residual transport plan tilde-T^m no independent evidence
    purpose: Added to the Sinkhorn OT plan to let the transport adapt to the downstream recommendation loss (Eq 12).
    No falsifiable prediction outside the training objective; its effect is measured only on recommendation metrics.

pith-pipeline@v1.3.0-alltime-deepseek · 16802 in / 17944 out tokens · 192043 ms · 2026-08-03T05:55:38.477592+00:00 · methodology

0 comments
read the original abstract

Integrating large language model (LLM) representations into multimodal recommendation has shown promise, yet a fundamental challenge remains largely overlooked: the semantic heterogeneity between generative LM representations and the ID-based collaborative signals that recommendation systems rely on. Naively injecting LM features without alignment degrades recommendation performance rather than improving it. To resolve this, we propose RecGOAT, a dual-granularity semantic alignment framework built on graph neural networks and optimal transport theory. RecGOAT first enriches collaborative semantics through multimodal attentive graphs that capture item-item, user-item, and user-user relationships, initializing user representations via LLM-inferred behavioral preferences. It then aligns LM-derived modality representations with recommendation IDs at two complementary granularities: (1) instance-level alignment via cross-modal contrastive learning (CMCL), which produces discriminative per-sample representations; and (2) distribution-level alignment via optimal adaptive transport (OAT), which minimizes the 1-Wasserstein distance between ID distributions and LLM semantics to produce a unified, consistently aligned feature space. Theoretically, we prove that the unified representation achieves strictly lower target error than any single-modality representation, with the gap bounded by the Wasserstein distance and the InfoNCE loss, providing rigorous guarantees for both alignment consistency and fusion comprehensiveness. Extensive experiments on three public benchmarks demonstrate state-of-the-art performance. Deployment on a large-scale online advertising platform further validates RecGOAT's industrial scalability. Our code is available at https://github.com/6lyc/RecGOAT-LLM4Rec.

Figures

Figures reproduced from arXiv: 2602.00682 by Chi Lu, Hengwei Ju, Kun Gai, Peng Jiang, Wei Yang, Yuecheng Li, Zeyu Song.

Figure 1
Figure 1. Figure 1: Performance comparison between LM representa [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overall framework of our RecGOAT. 3.1.1 Item-Item Multimodal Graph Representation Learning. To en￾hance high-order collaborative relationships between items, we con￾struct separate textual and visual modality graphs G 𝑚 = {I, E 𝑚, X𝑚 }, where𝑚 ∈ {𝑡, 𝑣 }. Inspired by FREEDOM [55], we adopt the K-nearest neighbors (KNN) algorithm to build frozen item-item graphs based on the initial LM-enhanced modal fea… view at source ↗
Figure 3
Figure 3. Figure 3: First, Figure (3a) demonstrates that the impact of differ [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. TimeMM: Time-as-Operator Spectral Filtering for Dynamic Multimodal Recommendation

    cs.IR 2026-04 unverdicted novelty 6.0

    TimeMM proposes a time-as-operator spectral filtering framework with adaptive mixing and modality routing to model non-stationary multimodal user preferences in recommendation systems.

  2. RecHarness: A Bandit-Routed Agentic Harness for Self-Evolving Recommender Systems

    cs.IR 2026-07 conditional novelty 5.0

    A bandit router allocates trials to human-defined edit directions while an LLM writes code within that direction, producing more stable recommender improvements and a positive online A/B result.

  3. Behavior-Guided Candidate Calibration for Multimodal Recommendation

    cs.IR 2026-05 unverdicted novelty 5.0

    Behavior-guided calibration converts co-user overlap into signed evidence applied only to multimodal recommender shortlists and yields consistent gains on Amazon Baby, Sports, and Electronics datasets.

  4. Taiji: Pareto Optimal Policy Optimization with Semantics-IDs Trade-off for Industrial LLM-Enhanced Recommendation

    cs.IR 2026-06 unverdicted novelty 4.0

    Taiji presents a LLM-as-Enhancer system with reverse-engineered CoT data generation and Pareto Optimal Policy Optimization (POPO) to trade off semantic and ID rewards, deployed at Kuaishou serving 400M daily users.

Reference graph

Works this paper leans on

62 extracted references · 5 linked inside Pith · cited by 4 Pith papers

  1. [1]

    Vineeta Anand and Ashish Kumar Maurya. 2025. A survey on recommender systems using graph neural network.ACM Transactions on Information Systems 43, 1 (2025), 1–49

    2025

  2. [2]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. Tallrec: An effective and efficient tuning framework to align large language model with recommendation. InProceedings of the 17th ACM conference on recommender systems. 1007–1014

    2023

  3. [3]

    Zongsheng Cao, Qianqian Xu, Zhiyong Yang, Yuan He, Xiaochun Cao, and Qingming Huang. 2022. OTKGE: Multi-modal Knowledge Graph Embeddings via Optimal Transport. InAdvances in Neural Information Processing Systems. https://openreview.net/forum?id=gbXqMdxsZIP Under Review, , Yuecheng Li, Hengwei Ju, Zeyu Song, Wei Yang, Chi Lu, Peng Jiang, and Kun Gai

    2022

  4. [4]

    Liqun Chen, Zhe Gan, Yu Cheng, Linjie Li, Lawrence Carin, and Jingjing Liu. 2020. Graph optimal transport for cross-domain alignment. InInternational Conference on Machine Learning. PMLR, 1542–1553

    2020

  5. [5]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. InInterna- tional conference on machine learning. PmLR, 1597–1607

    2020

  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. InProceedings of the 42nd international ACM SIGIR conference on research and development in information retrieval...

    2019

  7. [7]

    Nicolas Courty, Rémi Flamary, Amaury Habrard, and Alain Rakotomamonjy

  8. [8]

    Qiang Cui, Shu Wu, Qiang Liu, Wen Zhong, and Liang Wang. 2018. MV-RNN: A multi-view recurrent neural network for sequential recommendation.IEEE Transactions on Knowledge and Data Engineering32, 2 (2018), 317–331

    2018

  9. [9]

    Marco Cuturi. 2013. Sinkhorn distances: Lightspeed computation of optimal transport.Advances in neural information processing systems26 (2013)

    2013

  10. [10]

    Chen Gao, Yu Zheng, Nian Li, Yinfeng Li, Yingrong Qin, Jinghua Piao, Yuhan Quan, Jianxin Chang, Depeng Jin, Xiangnan He, et al. 2023. A survey of graph neural networks for recommender systems: Challenges, methods, and directions. ACM Transactions on Recommender Systems1, 1 (2023), 1–51

    2023

  11. [11]

    Albert Gu and Tri Dao. 2024. Mamba: Linear-time sequence modeling with selective state spaces. InFirst conference on language modeling

    2024

  12. [12]

    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

  13. [13]

    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

  14. [14]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. InProceedings of the 26th international conference on world wide web. 173–182

    2017

  15. [15]

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

  16. [16]

    Sein Kim, Hongseok Kang, Seungyoon Choi, Donghyun Kim, Minchul Yang, and Chanyoung Park. 2024. Large language models meet collaborative filtering: An efficient all-round llm-based recommender system. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1395–1406

    2024

  17. [17]

    Jiacheng Li, Ming Wang, Jin Li, Jinmiao Fu, Xin Shen, Jingbo Shang, and Julian McAuley. 2023. Text is all you need: Learning language representations for sequential recommendation. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1258–1267

    2023

  18. [18]

    Jin Li, Shoujin Wang, Qi Zhang, Shui Yu, and Fang Chen. 2025. Generating with fairness: A modality-diffused counterfactual framework for incomplete multimodal recommendations. InProceedings of the ACM on Web Conference 2025. 2787–2798

    2025

  19. [19]

    Yuecheng Li, Jialong Chen, Chuan Chen, Lei Yang, and Zibin Zheng. 2024. Con- trastive deep nonnegative matrix factorization for community detection. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Sig- nal Processing (ICASSP). IEEE, 6725–6729

    2024

  20. [20]

    Guojiao Lin, Meng Zhen, Dongjie Wang, Qingqing Long, Yuanchun Zhou, and Meng Xiao. 2024. GUME: Graphs and User Modalities Enhancement for Long- Tail Multimodal Recommendation. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management. 1400–1409

    2024

  21. [21]

    Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Hao Zhang, Yong Liu, Chuhan Wu, Xiangyang Li, Chenxu Zhu, et al . 2025. How can recommender systems benefit from large language models: A survey.ACM Transactions on Information Systems43, 2 (2025), 1–47

    2025

  22. [22]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual in- struction tuning.Advances in neural information processing systems36 (2023), 34892–34916

    2023

  23. [23]

    Yuqing Liu, Yu Wang, Lichao Sun, and Philip S Yu. 2024. Rec-gpt4v: Mul- timodal recommendation with large vision-language models.arXiv preprint arXiv:2402.08670(2024)

    Pith/arXiv arXiv 2024

  24. [24]

    Alejo Lopez-Avila and Jinhua Du. 2025. A Survey on Large Language Models in Multimodal Recommender Systems.arXiv preprint arXiv:2505.09777(2025)

    Pith/arXiv arXiv 2025

  25. [25]

    Daniele Malitesta, Giandomenico Cornacchia, Claudio Pomo, Felice Antonio Merra, Tommaso Di Noia, and Eugenio Di Sciascio. 2025. Formalizing multime- dia recommendation through multimodal deep learning.ACM Transactions on Recommender Systems3, 3 (2025), 1–33

    2025

  26. [26]

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

  27. [27]

    Gabriel Peyré. 2025. Optimal Transport for Machine Learners.arXiv preprint arXiv:2505.06589(2025)

    arXiv 2025

  28. [28]

    Gabriel Peyré, Marco Cuturi, et al. 2019. Computational optimal transport: With applications to data science.Foundations and Trends®in Machine Learning11, 5-6 (2019), 355–607

    2019

  29. [29]

    Claudio Pomo, Matteo Attimonelli, Danilo Danese, Fedelucio Narducci, and Tom- maso Di Noia. 2025. Do Recommender Systems Really Leverage Multimodal Content? A Comprehensive Analysis on Multimodal Representations for Recom- mendation. InProceedings of the 34th ACM International Conference on Information and Knowledge Management. 2377–2387

    2025

  30. [30]

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

  31. [31]

    Filippo Santambrogio. 2015. Optimal transport for applied mathematicians. (2015)

    2015

  32. [32]

    Richard Sinkhorn and Paul Knopp. 1967. Concerning nonnegative matrices and doubly stochastic matrices.Pacific J. Math.21, 2 (1967), 343–348

    1967

  33. [33]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.Advances in neural information processing systems30 (2017)

    2017

  34. [34]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. Graph Attention Networks. InInternational Con- ference on Learning Representations

    2018

  35. [35]

    Maolin Wang, Yutian Xiao, Binhao Wang, Sheng Zhang, Shanshan Ye, Wanyu Wang, Hongzhi Yin, Ruocheng Guo, and Zenglin Xu. 2025. FindRec: Stein-Guided Entropic Flow for Multi-Modal Sequential Recommendation. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 3008–3018

    2025

  36. [36]

    Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural graph collaborative filtering. InProceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval. 165–174

    2019

  37. [37]

    Yu Wang, Lei Sang, Yi Zhang, and Yiwen Zhang. 2025. Intent representation learning with large language model for recommendation. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1870–1879

    2025

  38. [38]

    Tianxin Wei, Bowen Jin, Ruirui Li, Hansi Zeng, Zhengyang Wang, Jianhui Sun, Qingyu Yin, Hanqing Lu, Suhang Wang, Jingrui He, and Xianfeng Tang. 2024. Towards Unified Multi-Modal Personalization: Large Vision-Language Models for Generative Recommendation and Beyond. InThe Twelfth International Conference on Learning Representations. https://openreview.net/...

    2024

  39. [39]

    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

  40. [40]

    Lianghao Xia, Chao Huang, Yong Xu, Jiashu Zhao, Dawei Yin, and Jimmy Huang

  41. [41]

    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

  42. [42]

    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)

    Pith/arXiv arXiv 2025

  43. [43]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report.arXiv preprint arXiv:2505.09388(2025)

    Pith/arXiv arXiv 2025

  44. [44]

    Wei Yang, Jie Yang, and Yuan Liu. 2023. Multimodal optimal transport knowledge distillation for cross-domain recommendation. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management. 2959–2968

    2023

  45. [45]

    Wei Yang and Qingchen Yang. 2024. Multimodal-aware Multi-intention Learning for Recommendation. InProceedings of the 32nd ACM International Conference on Multimedia. 5663–5672

    2024

  46. [46]

    Wei Yang, Rui Zhong, Yiqun Chen, Shixuan Li, Heng Ping, Chi Lu, and Peng Jiang. 2025. FITMM: Adaptive Frequency-Aware Multimodal Recommendation via Information-Theoretic Representation Learning. InProceedings of the 33rd ACM International Conference on Multimedia. 6193–6202

    2025

  47. [47]

    Zhengyi Yang, Jiancan Wu, Zhicai Wang, Xiang Wang, Yancheng Yuan, and Xiangnan He. 2023. Generate what you prefer: Reshaping sequential recommen- dation via guided diffusion.Advances in Neural Information Processing Systems 36 (2023), 24247–24261

    2023

  48. [48]

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

    2024

  49. [49]

    Zixuan Yi and Iadh Ounis. 2025. A Multi-modal Large Language Model with Graph-of-Thought for Effective Recommendation. InProceedings of the 2025 Con- ference of the Nations of the Americas Chapter of the Association for Computational RecGOAT Under Review, , Linguistics: Human Language Technologies (Volume 1: Long Papers). 1591–1606

    2025

  50. [50]

    Chao Zhang, Haoxin Zhang, Shiwei Wu, Di Wu, Tong Xu, Xiangyu Zhao, Yan Gao, Yao Hu, and Enhong Chen. 2025. NoteLLM-2: Multimodal Large Representation Models for Recommendation. In31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2025). Association for Computing Machinery, 2815–2826

    2025

  51. [51]

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

  52. [52]

    Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, et al. 2025. Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models.arXiv preprint arXiv:2506.05176(2025)

    Pith/arXiv arXiv 2025

  53. [53]

    Zihuai Zhao, Wenqi Fan, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Zhen Wen, Fei Wang, Xiangyu Zhao, Jiliang Tang, et al. 2024. Recommender systems in the era of large language models (llms).IEEE Transactions on Knowledge and Data Engineering36, 11 (2024), 6889–6907

    2024

  54. [54]

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

    2023

  55. [55]

    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

  56. [56]

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

    2023

  57. [2009]

    InProceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence

    BPR: Bayesian personalized ranking from implicit feedback. InProceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence. 452–461

  58. [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

  59. [2017]

    Joint distribution optimal transportation for domain adaptation.Advances in neural information processing systems30 (2017)

    2017

  60. [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

  61. [2022]

    InProceedings of the 45th International ACM SIGIR conference on research and development in information retrieval

    Hypergraph contrastive collaborative filtering. InProceedings of the 45th International ACM SIGIR conference on research and development in information retrieval. 70–79

  62. [2024]

    InProceedings of the 32nd ACM International Conference on Multimedia

    Diffmm: Multi-modal diffusion model for recommendation. InProceedings of the 32nd ACM International Conference on Multimedia. 7591–7599