Pith. sign in

REVIEW 3 major objections 4 minor 94 references

Is Personalized Modality Weighting Actually Personalized? A Controlled Audit of Per-User Weighting Claims in Multimodal Recommenders

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read This paper claims that per-user modality weighting—giving each user its own weights across visual, text, and audio channels—delivers no measurable ranking benefit over a single global weight once the user-weight binding is controlled with…

desk verdict A careful audit that cleanly separates 'the weight binds to the right user' from 'the weight beats a global one'; the negative utility result is solid as scoped, but the implant calibration never actually calibrates the utility contrast. read the letter →

arxiv 2608.05655 v1 pith:EN4JAEVK submitted 2026-08-06 cs.IR

Jingyuan Zheng , Xin Zhang , Yang Gu , Dongjing Wang , Yuxiang Wang , Xudong Shen , Haiping Zhang , Youhuizi Li
show 1 more author
Dongjin Yu
This is my paper · ORCID
classification cs.IR
keywords per-usermodalityweightingmultimodalrecommendationpersonalizationauditpermutationtestglobalweightbaselineidentifiabilityutilitydissociationimplicitfeedbackrecommendersystemsevaluation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Per-user modality weighting is a popular design in billion-user multimodal recommenders: each user gets a personal weight over visual, textual, or acoustic channels, on the claim that users differ in how much they trust each modality. This paper argues that published evaluations have never isolated a genuinely user-specific signal from a shared global weight plus the extra model capacity a per-user head adds. The audit reduces six published weighting methods onto one collaborative backbone and measures two contrasts: real-GM, how much a per-user weight beats a single global weight, and real-shuf, how much a user's learned weight is bound to that user. Across three short-video corpora and an e-commerce replication, a single global weight captures nearly all the content gain, and no per-user head wins consistently; several heads that look strongly personalized under the shuffle control simultaneously lose to the global weight. The paper proposes reporting both contrasts together as a minimum evidentiary standard for personalization claims.

What carries the argument

The carrying mechanism is a two-contrast audit principle applied to one shared collaborative backbone. Six weighting heads—a free per-user table, a bilinear attention gate, a meta-weight hypernetwork, a low-rank guided weight, and two decoupled variants—differ only in how the per-user modality weight is produced. The utility gap real-GM compares each trained head against a single global modality weight, and the identifiability gap real-shuf compares it against an evaluation-time permutation that swaps each user's weight with another user's from the same activity decile. A head counts as genuinely personalized only when real-GM is positive; real-shuf alone is necessary but not sufficient. The paper also uses gate-input decoupling as a causal probe to show that inflated shuffle gaps come from gates reading the shared collaborative embedding, not from a learned user-modality binding.

What would settle it

Run one of the audited heads in its original full architecture and training recipe on KuaiRand-27K and show it beats a single global weight by more than 1 percentage point on PairAcc across five seeds with FDR-corrected p < .05; that would contradict the claim that per-user weighting adds nothing beyond global weight plus capacity.

Watch

Extended reading notes

Core claim

The central claim is that, in the audited settings, the measurable value of per-user modality weighting is empirically bounded by a global modality weight plus model capacity, not by a user-specific modality-preference signal. On the main ranking probe (pairwise accuracy on held-out exposure pairs), a single global modality weight delivers +1.9, +3.6, and +3.5 percentage points over a no-modality baseline on the three short-video corpora, and per-user weights add no consistent utility: positive gaps are at most 0.9pp and flip sign across corpora and metrics, and all eighteen head-metric cells are negative on the Amazon-Baby replication. The paper further shows that a shuffle control can certify a weight as user-specific even when it is useless: the attention head on KuaiRand-27K posts an identifiability gap of +128% of the content gain while losing to the global weight by 0.7pp, and decoupling the gate input from the shared collaborative embedding collapses that gap to near zero. A monotone signal-implant dose-response confirms the null is not a measurement failure, because the harness would detect planted per-user preferences when they exist.

Load-bearing premise

The audit's reduction of six published weighting methods to reimplementations on one shared backbone is faithful enough that any real gains from the original methods would come through the weight-production mechanism alone; if the original systems' gains depend on interactions with their full architectures, the negative result may not transfer.

Editorial extensions

If this is right

  • Any per-user modality-weighting claim should report real-GM, not just real-shuf, with significance on at least two independent corpora.
  • A positive shuffle gap alone cannot certify useful personalization, because capacity absorbed by a gate reading the shared embedding can produce it.
  • The coldest-activity slice, the most plausible niche for per-user weighting, shows no benefit and is significantly negative on KuaiRand-27K.
  • The same utility-versus-identifiability distinction applies to any architecture that conditions parameters on a user embedding, not just modality weighting.
  • A capacity-matched global baseline loses to the simple global weight, so the global weight is near-optimal and not a strawman in these settings.

Reading between the lines

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

  • A two-contrast audit could be ported to other personalization claims, such as per-user temperature or per-user embeddings for fairness or safety, where a shuffle-only result is currently treated as proof of user-specific value.
  • The informational account suggests that scalar implicit feedback cannot credit which modality earned an interaction, so genuine per-user modality weights may require session-level traces, counterfactual exposures, or missing-modality natural experiments to learn.
  • The utility-identifiability plane plot is a generic diagnostic worth reusing: any point in the top-left region, highly identifiable yet with negative utility, is a candidate capacity artifact.
  • Dynamic or session-level per-user weights, which the paper excludes, could be tested with the same binding-permutation control applied at session granularity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes a two-contrast audit for per-user modality weights in multimodal recommender systems. Six weighting heads are reimplemented on a shared matrix-factorization backbone and compared against a global modality weight (real-GM) and against an eval-time shuffling of the user-weight binding (real-shuf). On three short-video corpora and one e-commerce corpus, a single global weight captures most of the content gain, per-user heads show no consistent utility advantage, and real-shuf can be large while real-GM is negative. The paper attributes the inflated real-shuf to gates reading the shared collaborative embedding, supports this with a decoupling intervention, and uses synthetic signal implants as a positive control. It concludes that personalization claims should report both contrasts.

Significance. This is a well-executed negative result with a reusable audit protocol. Its strengths are the controlled shared-backbone design, five-seed paired tests with FDR correction, a cross-domain replication, a LightGCN robustness check, a capacity-matched global baseline, and public release of code and per-seed results. The two-contrast criterion is a clear contribution. The main caveats are that the utility contrast is not directly calibrated by the implant experiment, and the shuffle implementation and the formal personalization criterion need tightening before the interpretive claims are fully supported.

major comments (3)
  1. [Instrument Calibration by Signal Implants; Results, Decoupling and Dose-Response Close the Loop] The implant calibration reports capture AUROC and the identifiability gap real-shuf, but not the utility gap real-GM that the paper defines as the personalization criterion. The statement in Results that 'the null utility result reflects an absent signal, not a blind instrument' does not follow from a monotone real-shuf response, because the paper's own dissociation (Table 4 versus Table 3: ATT on KuaiRand-27K has real-shuf +128% of content gain while real-GM is -0.69pp) shows real-shuf can be sensitive when real-GM is not. Appendix Table 12 likewise reports only AUROC. Please add the real-GM dose-response under implanted preferences, or explicitly limit the claim to 'no utility gain was detected in this harness' rather than 'no user-specific signal exists.'
  2. [Audit Design, Two Paired Contrasts; Results, Decoupling and Dose-Response Close the Loop] The formal criterion states that a head is genuinely personalized only when it beats GM on real-GM, but the discussion dismisses the surviving positive KuaiRand meta-weight cells (Table 3: MWNd +0.26 PairAcc, +0.16 NDCG, +0.32 Recall, all FDR-significant) because real-shuf is near zero, calling the gain 'capacity, not personalization.' This introduces a second condition that is not part of the stated criterion. If real-GM>0 is intended as necessary but not sufficient, the full condition should be defined; if it is sufficient, those cells are a counterexample to the summary claim that no per-user head adds utility. The paper should reconcile the criterion with this case.
  3. [Audit Design, Two Paired Contrasts] The abstract and introduction describe real-shuf as an 'eval-time permutation of the user-weight binding,' but the implementation described in Audit Design replaces each user's weight with a different user drawn from the same activity decile, averaged over five draws. If these are independent draws with replacement, the multiset of weights is not preserved and the gap can include a distribution-shift component; if they are true permutations, the wording should say so. Please clarify the sampling scheme and, if it is with replacement, either switch to a random permutation within each decile or justify the with-replacement design and rename the contrast.
minor comments (4)
  1. [Figure 1, panel (c)] The phrase 'costs +4.64 percentage points' is confusing because Table 4 reports the gap as +4.64pp; a drop should be described as a drop of 4.64pp or as -4.64pp to avoid sign ambiguity.
  2. [Tables 2, 3, 4, and 10] Table 2 uses p-values while Tables 3, 4, and 10 use FDR q-values; the captions should state this explicitly in each table so that readers do not compare stars across tables as if they were the same quantity.
  3. [Related Work, Per-User Modality Weighting] The reference to Shu et al. (2019) for MWN is to a sample-weighting hypernetwork rather than a modality-weighting method; the text should clarify that the head is inspired by, rather than identical to, the cited method.
  4. [Results, Decoupling and Dose-Response Close the Loop] The statement that a gain surviving weight swaps is 'capacity, not personalization' is not directly tested by the capacity-matched GMc baseline, because GMc allocates capacity to item content rather than user identity; a sentence acknowledging this distinction would prevent overinterpretation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the audit is an externally benchmarked empirical comparison with an independent positive control.

full rationale

The paper's central claims are empirical contrasts computed on public datasets against a fixed global baseline, not conclusions derived from their own definitions. The utility gap real-GM is defined as Perf(hreal) - Perf(GM) and is measured directly; no parameter fitted to reproduce that gap is then presented as a prediction. The implant calibration is a positive control for instrument sensitivity: it tracks capture AUROC and the identifiability gap under synthetic implants, and the null utility result is reported as a direct measurement rather than as an output of that calibration. Even if the calibration omits a direct real-GM dose-response, this is a validity limitation of the control, not a definitional dependency. The paper also explicitly acknowledges the reimplementation-fidelity assumption and proxies, which are honest scope limits rather than circular moves. References to prior work are external datasets and methods; there is no load-bearing self-citation chain or imported uniqueness theorem. The decoupling analysis is an experimental intervention, not an ansatz smuggled through citation. Because the main findings are externally falsifiable comparisons with released code and per-seed results, no step reduces by construction to its own input.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a controlled empirical comparison. The only fitted quantities are the model's own weights and the standard baselines; no hidden constants are introduced. Stated hyperparameters are fixed across heads to avoid per-head tuning.

free parameters (3)
  • Global modality weight w_m for GM baseline = learned per corpus
    The global weight is the non-personalized comparator; it is fit to each corpus and is the object against which all per-user heads are judged. It is a stated experimental baseline, not a hidden assumption.
  • Global content scale c = learned single scalar
    Scales modality cosine similarity into a bounded range and is shared across all heads; learned from data. It is part of the backbone and not per-user.
  • Low-rank guided weight rank = 8
    Hand-chosen rank for the LRG head; one of several fixed hyperparameters, stated in Table 5. It is a design choice but not central to the conclusion.
assumptions (5)
  • standard math Matrix factorization and BPR pairwise objective are a valid testbed for comparing weighting mechanisms.
    The shared backbone is MF with BPR objective, a standard collaborative filtering setup; cited to Rendle et al. 2009 and Koren et al. 2009.
  • domain assumption Watch ratio is a valid implicit preference proxy and duration is excluded from modality channels.
    Dataset cards state watch ratio as implicit feedback and duration excluded to prevent coupling with label; the audit's outcomes depend on this proxy.
  • domain assumption The six reimplementations faithfully capture the weight-production mechanisms of the cited per-user weighting family.
    The reduction to one backbone assumes the family shares its behavior in the head; stated in Audit Design and Limitations.
  • domain assumption Eval-time within-activity-decile permutation of user weights isolates the user-modality binding from activity confounds.
    Audit Design constrains shuffles to the same activity decile and applies them after training; this assumption defines the identifiability gap.
  • domain assumption Implanted synthetic modality preferences behave like natural user-specific preferences for calibrating the detector.
    Instrument Calibration plants a preferred modality with strength alpha and uses capture AUROC to conclude the null is not a measurement failure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Is Personalized Modality Weighting Actually Personalized? A Controlled Audit of Per-User Weighting Claims in Multimodal Recommenders." pith.science (2026). https://pith.science/paper/EN4JAEVK

@misc{pith2026260805655,
  author       = {Pith},
  title        = {Pith review of: Is Personalized Modality Weighting Actually Personalized? A Controlled Audit of Per-User Weighting Claims in Multimodal Recommenders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EN4JAEVK}},
  note         = {Machine review of arXiv:2608.05655}
}
read the original abstract

Per-user modality weighting is deployed at billion-user scale in multimodal recommenders, through user modality-strength vectors, attention gates, meta-weight hypernetworks, and low-rank guided weights, each claiming a ranking gain from user-specific modality preference. Yet, to our knowledge, prior evaluations do not isolate a genuinely user-specific signal from a global modality weight plus model capacity. We audit this family with a two-contrast audit principle, reducing six implementations onto one shared collaborative backbone and measuring a utility gap (real-GM) against a single global modality weight and an identifiability gap (real-shuf) against an eval-time permutation of the user-weight binding. Across three independent short-video corpora, a single global weight already delivers nearly all of the content gain (+1.9/+3.6/+3.5pp over a no-modality baseline, p < .001). Making the weight per-user adds no consistent utility: no implementation wins on all corpora and metrics, and the few positive gaps are small (<=0.9pp) and flip. The shuffle control is necessary but not sufficient, since real-shuf reaches +128% of the content gain for heads that simultaneously lose to the global weight. We trace this dissociation to gates reading the shared collaborative embedding: decoupling the gate input collapses the inflated real-shuf to near zero while the utility conclusion stands. A monotone signal-implant dose-response (capture AUROC rising from 0.57 to 0.89 and from 0.64 to 1.00) verifies the harness would detect user-specific structure if present, and every finding replicates on a fourth, cross-domain e-commerce corpus. We propose reporting real-GM alongside real-shuf as a minimum evidentiary standard for personalization claims.

Figures

Figures reproduced from arXiv: 2608.05655 by the authors.

Figure 1
Figure 1. Different per-user modality weights do not imply useful modality personalization. The three panels ask in turn whether [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Utility versus identifiability for six heads on three [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. (a) Decoupling the gate input from the shared col [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

94 extracted references · 77 canonical work pages

  1. [1]

    Adebayo, J.; Gilmer, J.; Muelly, M.; Goodfellow, I.; Hardt, M.; and Kim, B. 2018. Sanity Checks for Saliency Maps. In Advances in Neural Information Processing Systems (NeurIPS), volume 31, 9525--9536

    2018

  2. [2]

    W.; Malitesta, D.; Pomo, C.; Bellog \' n, A.; Di Sciascio, E.; and Di Noia, T

    Anelli, V. W.; Malitesta, D.; Pomo, C.; Bellog \' n, A.; Di Sciascio, E.; and Di Noia, T. 2023. Challenging the Myth of Graph Collaborative Filtering: A Reasoned and Reproducibility-driven Analysis. In Proceedings of the 17th ACM Conference on Recommender Systems (RecSys '23), 350--361. ACM

    2023

  3. [3]

    Breiman, L. 2001. Random Forests. Machine Learning, 45(1): 5--32

    2001

  4. [4]

    Chen, Y.; Sun, F.; Jiang, Y.; Cao, Y.; Zhang, Y.; Lin, Z.; Li, Z.; Gai, K.; and Zhang, G. 2024. A Multi-modal Modeling Framework for Cold-start Short-video Recommendation. In Proceedings of the 18th ACM Conference on Recommender Systems (RecSys '24), 391--400. ACM

    2024

  5. [5]

    Cremonesi, P.; Koren, Y.; and Turrin, R. 2010. Performance of Recommender Algorithms on Top-N Recommendation Tasks. In Proceedings of the 4th ACM Conference on Recommender Systems (RecSys '10), 39--46. ACM

    2010

  6. [7]

    Ferrari Dacrema, M.; Boglio, S.; Cremonesi, P.; and Jannach, D. 2021. A Troubling Analysis of Reproducibility and Progress in Recommender Systems Research. ACM Transactions on Information Systems, 39(2): 20:1--20:49

    2021

  7. [8]

    Ferrari Dacrema, M.; Cremonesi, P.; and Jannach, D. 2019. Are We Really Making Much Progress? A Worrying Analysis of Recent Neural Recommendation Approaches. In Proceedings of the 13th ACM Conference on Recommender Systems (RecSys '19), 101--109. ACM

    2019

  8. [9]

    Fisher, A.; Rudin, C.; and Dominici, F. 2019. All Models are Wrong, but Many are Useful: Learning a Variable's Importance by Studying an Entire Class of Prediction Models Simultaneously. Journal of Machine Learning Research, 20(177): 1--81

    2019

Show all 94 references
  1. [10]

    Gao, C.; Li, S.; Zhang, Y.; Chen, J.; Li, B.; Lei, W.; Jiang, P.; and He, X. 2022. KuaiRand: An Unbiased Sequential Recommendation Dataset with Randomly Exposed Videos. In Proceedings of the 31st ACM International Conference on Information and Knowledge Management (CIKM '22), ...

    2022

  2. [11]

    Guo, Z.; Li, J.; Li, G.; Wang, C.; Shi, S.; and Ruan, B. 2024. LGMRec: Local and Global Graph Learning for Multimodal Recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 8454--8462

    2024

  3. [12]

    He, R.; and McAuley, J. 2016. VBPR: Visual Bayesian Personalized Ranking from Implicit Feedback. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 30, 144--150

    2016

  4. [13]

    He, X.; Deng, K.; Wang, X.; Li, Y.; Zhang, Y.; and Wang, M. 2020. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '20), 63...

    2020

  5. [14]

    Hooker, G.; Mentch, L.; and Zhou, S. 2021. Unrestricted Permutation Forces Extrapolation: Variable Importance Requires at Least One More Model, or There Is No Free Variable Importance. Statistics and Computing, 31(6): 82

    2021

  6. [15]

    Hooker, S.; Erhan, D.; Kindermans, P.-J.; and Kim, B. 2019. A Benchmark for Interpretability Methods in Deep Neural Networks. In Advances in Neural Information Processing Systems (NeurIPS), volume 32, 9734--9745

    2019

  7. [16]

    Hu, Y.; Koren, Y.; and Volinsky, C. 2008. Collaborative Filtering for Implicit Feedback Datasets. In Proceedings of the 8th IEEE International Conference on Data Mining (ICDM '08), 263--272. IEEE

    2008

  8. [17]

    Koren, Y.; Bell, R.; and Volinsky, C. 2009. Matrix Factorization Techniques for Recommender Systems. Computer, 42(8): 30--37

    2009

  9. [18]

    Krichene, W.; and Rendle, S. 2020. On Sampled Metrics for Item Recommendation. In Proceedings of the 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD '20). ACM

    2020

  10. [19]

    Li, H.; Du, H.; Li, Y.; Fu, J.; Li, C.; Zhuang, Z.; Li, J.; and Ni, Y. 2025. Teach Me How to Denoise: A Universal Framework for Denoising Multi-modal Recommender Systems via Guided Calibration. In Proceedings of the 18th ACM International Conference on Web Search and Data Mini...

    2025

  11. [20]

    Lin, J. 2019. The Neural Hype and Comparisons Against Weak Baselines. SIGIR Forum, 52(2): 40--51

    2019

  12. [21]

    C.; and Steinhardt, J

    Lipton, Z. C.; and Steinhardt, J. 2019. Research for Practice: Troubling Trends in Machine-Learning Scholarship. Communications of the ACM, 62(6): 45--53

    2019

  13. [22]

    McAuley, J.; Targett, C.; Shi, Q.; and van den Hengel, A. 2015. Image-Based Recommendations on Styles and Substitutes. In Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '15), 43--52. ACM

    2015

  14. [23]

    Musgrave, K.; Belongie, S.; and Lim, S.-N. 2020. A Metric Learning Reality Check. In Computer Vision -- ECCV 2020, volume 12370 of Lecture Notes in Computer Science, 681--699. Springer

    2020

  15. [24]

    Ni, Y.; Bi, L.; Han, J.; Zhao, G.; Wang, Y.; Li, S.; Zhang, J.; Mao, J.; Wei, Y.; Ma, W.; Ye, Y.; and Zhang, M. 2023. MicroLens : A Relevance-based Micro-video Recommendation Dataset with Natural Language Explanations. In Advances in Neural Information Processing Systems (Neur...

    2023

  16. [25]

    Ojala, M.; and Garriga, G. C. 2010. Permutation Tests for Studying Classifier Performance. Journal of Machine Learning Research, 11(62): 1833--1863

    2010

  17. [26]

    Ong, R.; and Khong, A. W. H. 2025. Spectrum-based Modality Representation Fusion Graph Convolutional Network for Multimodal Recommendation. In Proceedings of the 18th ACM International Conference on Web Search and Data Mining (WSDM '25), 773--781. ACM

    2025

  18. [27]

    Qi, F.; Zhang, D.; Lin, X.; Su, X.; Zhu, Z.; Wang, X.; and Li, K. 2025. Seeing Beyond Noise: Joint Graph Structure Evaluation and Denoising for Multimodal Recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 12461--12469

    2025

  19. [28]

    Rendle, S.; Freudenthaler, C.; Gantner, Z.; and Schmidt-Thieme, L. 2009. BPR: Bayesian Personalized Ranking from Implicit Feedback. In Proceedings of the 25th Conference on Uncertainty in Artificial Intelligence (UAI '09), 452--461

    2009

  20. [29]

    Rendle, S.; Krichene, W.; Zhang, L.; and Anderson, J. R. 2020. Neural Collaborative Filtering vs. Matrix Factorization Revisited. In Proceedings of the 14th ACM Conference on Recommender Systems (RecSys '20), 240--248. ACM

    2020

  21. [30]

    Sculley, D.; Snoek, J.; Wiltschko, A.; and Rahimi, A. 2018. Winner's Curse? On Pace, Progress, and Empirical Rigor. In 6th International Conference on Learning Representations (ICLR '18), Workshop Track

    2018

  22. [31]

    Shang, Y.; Gao, C.; Li, N.; and Li, Y. 2025. A Large-scale Dataset with Behavior, Attributes, and Content of Mobile Short-video Platform. In Companion Proceedings of the ACM Web Conference 2025 (WWW '25 Companion), 793--796. ACM

    2025

  23. [32]

    Shu, J.; Xie, Q.; Yi, L.; Zhao, Q.; Zhou, S.; Xu, Z.; and Meng, D. 2019. Meta-Weight-Net: Learning an Explicit Mapping for Sample Weighting. In Advances in Neural Information Processing Systems (NeurIPS), volume 32

    2019

  24. [33]

    Strobl, C.; Boulesteix, A.-L.; Kneib, T.; Augustin, T.; and Zeileis, A. 2008. Conditional Variable Importance for Random Forests. BMC Bioinformatics, 9: 307

    2008

  25. [34]

    Tao, Z.; Liu, X.; Xia, Y.; Wang, X.; Yang, L.; Huang, X.; and Chua, T.-S. 2023. Self-Supervised Learning for Multimedia Recommendation. IEEE Transactions on Multimedia, 25: 5107--5116

    2023

  26. [35]

    Tao, Z.; Wei, Y.; Wang, X.; He, X.; Huang, X.; and Chua, T.-S. 2020. MGAT : Multimodal Graph Attention Network for Recommendation. Information Processing & Management, 57(5): 102277

    2020

  27. [36]

    Wang, Q.; Wei, Y.; Yin, J.; Wu, J.; Song, X.; and Nie, L. 2023. DualGNN: Dual Graph Neural Network for Multimedia Recommendation. IEEE Transactions on Multimedia, 25: 1074--1084

    2023

  28. [37]

    Wang, X.; He, X.; Wang, M.; Feng, F.; and Chua, T.-S. 2019. Neural Graph Collaborative Filtering. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '19), 165--174. ACM

    2019

  29. [38]

    Wei, W.; Huang, C.; Xia, L.; and Zhang, C. 2023. Multi-Modal Self-Supervised Learning for Recommendation. In Proceedings of the ACM Web Conference 2023 (WWW '23), 790--800. ACM

    2023

  30. [39]

    Wei, Y.; Wang, X.; Nie, L.; He, X.; and Chua, T.-S. 2020. Graph-Refined Convolutional Network for Multimedia Recommendation with Implicit Feedback. In Proceedings of the 28th ACM International Conference on Multimedia (MM '20), 3541--3549. ACM

    2020

  31. [40]

    Wei, Y.; Wang, X.; Nie, L.; He, X.; Hong, R.; and Chua, T.-S. 2019. MMGCN: Multi-modal Graph Convolution Network for Personalized Recommendation of Micro-video. In Proceedings of the 27th ACM International Conference on Multimedia (MM '19), 1437--1445. ACM

    2019

  32. [41]

    Xu, J.; Chen, Z.; Yang, S.; Li, J.; Wang, H.; and Ngai, E. C.-H. 2025. MENTOR: Multi-level Self-supervised Learning for Multimodal Recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 12908--12917

    2025

  33. [42]

    Yu, P.; Tan, Z.; Lu, G.; and Bao, B.-K. 2023. Multi-View Graph Convolutional Network for Multimedia Recommendation. In Proceedings of the 31st ACM International Conference on Multimedia (MM '23), 6576--6585. ACM

    2023

  34. [43]

    Zhang, J.; Zhu, Y.; Liu, Q.; Wu, S.; Wang, S.; and Wang, L. 2021. Mining Latent Structures for Multimodal Recommendation. In Proceedings of the 29th ACM International Conference on Multimedia (MM '21), 3872--3880. ACM

    2021

  35. [44]

    Zhang, J.; Zhu, Y.; Liu, Q.; Zhang, M.; Wu, S.; and Wang, L. 2023. Latent Structure Mining with Contrastive Modality Fusion for Multimedia Recommendation. IEEE Transactions on Knowledge and Data Engineering, 35(9): 9154--9167

    2023

  36. [46]

    Zhou, H.; Zhou, X.; Zhang, L.; and Shen, Z. 2023 b . Enhancing Dyadic Relations with Homogeneous Graphs for Multimodal Recommendation. In Proceedings of the 26th European Conference on Artificial Intelligence (ECAI '23), volume 372 of Frontiers in Artificial Intelligence and A...

    2023

  37. [47]

    Zhou, X.; and Shen, Z. 2023. A Tale of Two Graphs: Freezing and Denoising Graph Structures for Multimodal Recommendation. In Proceedings of the 31st ACM International Conference on Multimedia (MM '23), 935--943. ACM

    2023

  38. [48]

    Zhou, X.; Zhou, H.; Liu, Y.; Zeng, Z.; Miao, C.; Wang, P.; You, Y.; and Jiang, F. 2023 c . Bootstrap Latent Representations for Multi-modal Recommendation. In Proceedings of the ACM Web Conference 2023 (WWW '23), 845--854. ACM

    2023

  39. [49]

    Proceedings of the 18th ACM Conference on Recommender Systems (RecSys '24) , year =

    Chen, Ye and Sun, Fei and Jiang, Yifan and Cao, Yuxi and Zhang, Yanbin and Lin, Zhiyong and Li, Zhen and Gai, Kun and Zhang, Guorui , title =. Proceedings of the 18th ACM Conference on Recommender Systems (RecSys '24) , year =

  40. [50]

    arXiv preprint arXiv:2504.16524 , year =

    Dong, Xue and Song, Xuemeng and Zheng, Na and Zhao, Sicheng and Ding, Guiguang , title =. arXiv preprint arXiv:2504.16524 , year =

    arXiv

  41. [51]

    Ong, Rongqing and Khong, Andy W. H. , title =. Proceedings of the 18th ACM International Conference on Web Search and Data Mining (WSDM '25) , year =

  42. [52]

    , title =

    Xu, Jinfeng and Chen, Zheyu and Yang, Shuo and Li, Jinze and Wang, Hewei and Ngai, Edith C.-H. , title =. Proceedings of the AAAI Conference on Artificial Intelligence , year =

  43. [53]

    Proceedings of the AAAI Conference on Artificial Intelligence , year =

    Qi, Feng and Zhang, Debing and Lin, Xiao and Su, Xiaowen and Zhu, Zhiwei and Wang, Xin and Li, Ke , title =. Proceedings of the AAAI Conference on Artificial Intelligence , year =

  44. [54]

    Proceedings of the 18th ACM International Conference on Web Search and Data Mining (WSDM '25) , year =

    Li, Hongji and Du, Hanwen and Li, Youhua and Fu, Junchen and Li, Chunxiao and Zhuang, Ziyi and Li, Jiakang and Ni, Yongxin , title =. Proceedings of the 18th ACM International Conference on Web Search and Data Mining (WSDM '25) , year =

  45. [55]

    Proceedings of the 25th Conference on Uncertainty in Artificial Intelligence (UAI '09) , year =

    Rendle, Steffen and Freudenthaler, Christoph and Gantner, Zeno and Schmidt-Thieme, Lars , title =. Proceedings of the 25th Conference on Uncertainty in Artificial Intelligence (UAI '09) , year =

  46. [56]

    Proceedings of the AAAI Conference on Artificial Intelligence , year =

    He, Ruining and McAuley, Julian , title =. Proceedings of the AAAI Conference on Artificial Intelligence , year =

  47. [57]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Shu, Jun and Xie, Qi and Yi, Lixuan and Zhao, Qian and Zhou, Sanping and Xu, Zongben and Meng, Deyu , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  48. [58]

    Journal of Machine Learning Research , year =

    Fisher, Aaron and Rudin, Cynthia and Dominici, Francesca , title =. Journal of Machine Learning Research , year =

  49. [59]

    SIGIR Forum , volume =

    Lin, Jimmy , title =. SIGIR Forum , volume =. 2019 , doi =

    2019

  50. [60]

    Proceedings of the 13th ACM Conference on Recommender Systems (RecSys '19) , year =

    Ferrari Dacrema, Maurizio and Cremonesi, Paolo and Jannach, Dietmar , title =. Proceedings of the 13th ACM Conference on Recommender Systems (RecSys '19) , year =

  51. [61]

    Proceedings of the 31st ACM International Conference on Multimedia (MM '23) , year =

    Zhou, Xin and Shen, Zhiqi , title =. Proceedings of the 31st ACM International Conference on Multimedia (MM '23) , year =

  52. [62]

    Proceedings of the ACM Web Conference 2023 (WWW '23) , year =

    Zhou, Xin and Zhou, Hongyu and Liu, Yong and Zeng, Zhiwei and Miao, Chunyan and Wang, Pengwei and You, Yuan and Jiang, Feijun , title =. Proceedings of the ACM Web Conference 2023 (WWW '23) , year =

    2023

  53. [63]

    ACM Transactions on Information Systems , year =

    Ferrari Dacrema, Maurizio and Boglio, Simone and Cremonesi, Paolo and Jannach, Dietmar , title =. ACM Transactions on Information Systems , year =

  54. [64]

    Proceedings of the 27th ACM International Conference on Multimedia (MM '19) , year =

    Wei, Yinwei and Wang, Xiang and Nie, Liqiang and He, Xiangnan and Hong, Richang and Chua, Tat-Seng , title =. Proceedings of the 27th ACM International Conference on Multimedia (MM '19) , year =

  55. [65]

    Proceedings of the 28th ACM International Conference on Multimedia (MM '20) , year =

    Wei, Yinwei and Wang, Xiang and Nie, Liqiang and He, Xiangnan and Chua, Tat-Seng , title =. Proceedings of the 28th ACM International Conference on Multimedia (MM '20) , year =

  56. [66]

    IEEE Transactions on Multimedia , volume =

    Wang, Qifan and Wei, Yinwei and Yin, Jianhua and Wu, Jianlong and Song, Xuemeng and Nie, Liqiang , title =. IEEE Transactions on Multimedia , volume =. 2023 , doi =

    2023

  57. [67]

    Proceedings of the AAAI Conference on Artificial Intelligence , year =

    Guo, Zhiqiang and Li, Jianjun and Li, Guohui and Wang, Chaoyang and Shi, Si and Ruan, Bin , title =. Proceedings of the AAAI Conference on Artificial Intelligence , year =

  58. [68]

    IEEE Transactions on Multimedia , volume =

    Tao, Zhulin and Liu, Xiaohao and Xia, Yewei and Wang, Xiang and Yang, Lifang and Huang, Xianglin and Chua, Tat-Seng , title =. IEEE Transactions on Multimedia , volume =. 2023 , doi =

    2023

  59. [69]

    Proceedings of the ACM Web Conference 2023 (WWW '23) , year =

    Wei, Wei and Huang, Chao and Xia, Lianghao and Zhang, Chuxu , title =. Proceedings of the ACM Web Conference 2023 (WWW '23) , year =

    2023

  60. [70]

    arXiv preprint arXiv:2302.04473 , year =

    Zhou, Hongyu and Zhou, Xin and Zeng, Zhiwei and Zhang, Lingzi and Shen, Zhiqi , title =. arXiv preprint arXiv:2302.04473 , year =

    arXiv

  61. [71]

    Computer , volume =

    Koren, Yehuda and Bell, Robert and Volinsky, Chris , title =. Computer , volume =. 2009 , doi =

    2009

  62. [72]

    Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '19) , year =

    Wang, Xiang and He, Xiangnan and Wang, Meng and Feng, Fuli and Chua, Tat-Seng , title =. Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '19) , year =

  63. [73]

    Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '20) , year =

    He, Xiangnan and Deng, Kuan and Wang, Xiang and Li, Yan and Zhang, Yongdong and Wang, Meng , title =. Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '20) , year =

  64. [74]

    Machine Learning , volume =

    Breiman, Leo , title =. Machine Learning , volume =. 2001 , doi =

    2001

  65. [75]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Adebayo, Julius and Gilmer, Justin and Muelly, Michael and Goodfellow, Ian and Hardt, Moritz and Kim, Been , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  66. [76]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Hooker, Sara and Erhan, Dumitru and Kindermans, Pieter-Jan and Kim, Been , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  67. [77]

    , title =

    Rendle, Steffen and Krichene, Walid and Zhang, Li and Anderson, John R. , title =. Proceedings of the 14th ACM Conference on Recommender Systems (RecSys '20) , year =

  68. [78]

    Proceedings of the 4th ACM Conference on Recommender Systems (RecSys '10) , year =

    Cremonesi, Paolo and Koren, Yehuda and Turrin, Roberto , title =. Proceedings of the 4th ACM Conference on Recommender Systems (RecSys '10) , year =

  69. [79]

    Challenging the Myth of Graph Collaborative Filtering: A Reasoned and Reproducibility-driven Analysis , booktitle =

    Anelli, Vito Walter and Malitesta, Daniele and Pomo, Claudio and Bellog. Challenging the Myth of Graph Collaborative Filtering: A Reasoned and Reproducibility-driven Analysis , booktitle =. 2023 , pages =

    2023

  70. [80]

    Computer Vision -- ECCV 2020 , series =

    Musgrave, Kevin and Belongie, Serge and Lim, Ser-Nam , title =. Computer Vision -- ECCV 2020 , series =. 2020 , publisher =

    2020

  71. [81]

    and Steinhardt, Jacob , title =

    Lipton, Zachary C. and Steinhardt, Jacob , title =. Communications of the ACM , volume =. 2019 , doi =

    2019

  72. [82]

    Proceedings of the 31st ACM International Conference on Information and Knowledge Management (CIKM '22) , year =

    Gao, Chongming and Li, Shijun and Zhang, Yuan and Chen, Jiawei and Li, Biao and Lei, Wenqiang and Jiang, Peng and He, Xiangnan , title =. Proceedings of the 31st ACM International Conference on Information and Knowledge Management (CIKM '22) , year =

  73. [83]

    Companion Proceedings of the ACM Web Conference 2025 (WWW '25 Companion) , year =

    Shang, Yu and Gao, Chen and Li, Nian and Li, Yong , title =. Companion Proceedings of the ACM Web Conference 2025 (WWW '25 Companion) , year =

    2025

  74. [84]

    Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track , year =

    Ni, Yinsheng and Bi, Ling and Han, Jiangong and Zhao, Guohao and Wang, Yong and Li, Shanghang and Zhang, Jieming and Mao, Jiaxin and Wei, Yujia and Ma, Weizhi and Ye, Yunshan and Zhang, Min , title =. Advances in Neural Information Processing Systems (NeurIPS) Datasets and Ben...

  75. [85]

    Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '15) , year =

    McAuley, Julian and Targett, Christopher and Shi, Qinfeng and van den Hengel, Anton , title =. Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '15) , year =

  76. [86]

    , title =

    Ojala, Markus and Garriga, Gemma C. , title =. Journal of Machine Learning Research , year =

  77. [87]

    BMC Bioinformatics , year =

    Strobl, Carolin and Boulesteix, Anne-Laure and Kneib, Thomas and Augustin, Thomas and Zeileis, Achim , title =. BMC Bioinformatics , year =

  78. [88]

    Statistics and Computing , year =

    Hooker, Giles and Mentch, Lucas and Zhou, Siyu , title =. Statistics and Computing , year =

  79. [89]

    Proceedings of the 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD '20) , year =

    Krichene, Walid and Rendle, Steffen , title =. Proceedings of the 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD '20) , year =

  80. [90]

    Proceedings of the 8th IEEE International Conference on Data Mining (ICDM '08) , year =

    Hu, Yifan and Koren, Yehuda and Volinsky, Chris , title =. Proceedings of the 8th IEEE International Conference on Data Mining (ICDM '08) , year =

  81. [91]

    Proceedings of the 29th ACM International Conference on Multimedia (MM '21) , year =

    Zhang, Jinghao and Zhu, Yanqiao and Liu, Qiang and Wu, Shu and Wang, Shuhui and Wang, Liang , title =. Proceedings of the 29th ACM International Conference on Multimedia (MM '21) , year =

  82. [92]

    Proceedings of the 31st ACM International Conference on Multimedia (MM '23) , year =

    Yu, Penghang and Tan, Zhiyi and Lu, Guanming and Bao, Bing-Kun , title =. Proceedings of the 31st ACM International Conference on Multimedia (MM '23) , year =

  83. [93]

    Information Processing & Management , volume =

    Tao, Zhulin and Wei, Yinwei and Wang, Xiang and He, Xiangnan and Huang, Xianglin and Chua, Tat-Seng , title =. Information Processing & Management , volume =. 2020 , doi =

    2020

  84. [94]

    IEEE Transactions on Knowledge and Data Engineering , volume =

    Zhang, Jinghao and Zhu, Yanqiao and Liu, Qiang and Zhang, Mengqi and Wu, Shu and Wang, Liang , title =. IEEE Transactions on Knowledge and Data Engineering , volume =. 2023 , doi =

    2023

  85. [95]

    Proceedings of the 26th European Conference on Artificial Intelligence (ECAI '23) , series =

    Zhou, Hongyu and Zhou, Xin and Zhang, Lingzi and Shen, Zhiqi , title =. Proceedings of the 26th European Conference on Artificial Intelligence (ECAI '23) , series =. 2023 , publisher =

    2023

  86. [96]

    and Snoek, Jasper and Wiltschko, Alex and Rahimi, Ali , title =

    Sculley, D. and Snoek, Jasper and Wiltschko, Alex and Rahimi, Ali , title =. 6th International Conference on Learning Representations (ICLR '18), Workshop Track , year =

Pith tools

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