Pith. sign in

REVIEW 3 major objections 5 minor 55 references

This paper claims that a two-level denoising framework—graph-based feature denoising and FFT-based sequence denoising—consistently outperforms state-of-the-art multi-modal sequential recommenders, with Recall@20 gains up to 19.33%.

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

2026-08-01 14:18 UTC pith:IHNT5GKQ

load-bearing objection A practically useful integration of known denoising ideas with strong reported gains, but the conflicting numbers between Table 2 and Table 4 need to be resolved before anyone can trust the headline results. the 3 major comments →

arxiv 2607.18786 v1 pith:IHNT5GKQ submitted 2026-07-21 cs.IR

Beyond Noisy Signals: Dual-Level Denoising for Multi-modal Sequential Recommendation

Jie Luo , Qi Jin , Xinming Zhang This is my paper
classification cs.IR
keywords sequential recommendationmulti-modal recommendationdenoisingFast Fourier Transformgraph signal processingLaplacian smoothingcontrastive learningfrequency-domain filtering
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 paper sets out to solve what it calls the 'Dual-Noise Dilemma' in multi-modal sequential recommendation: pre-trained text and image features contain task-irrelevant redundancy, and user interaction sequences contain accidental clicks and other spurious signals. It proposes DDMSR, which denoises features by Laplacian smoothing on item-item semantic graphs (a structural low-pass filter) and denoises sequences by learnable filtering in the Fourier domain, with a contrastive objective aligning text and image embeddings. On four public datasets—Beauty, Sports, Toys, and MicroLens—DDMSR reports consistent state-of-the-art results, with Recall@20 gains up to +19.33% (Sports) and NDCG@20 gains up to +10.68% (Beauty) over the strongest baseline. If right, this means a lightweight, non-generative alternative to diffusion-based denoising for recommendation, with long-tail items also benefiting through semantic neighbor aggregation.

Core claim

The central discovery asserted by the paper is that noise in multi-modal sequential recommendation can be attacked at two independent levels using spectral reasoning. Features extracted by frozen pre-trained encoders live on a noisy manifold: semantically similar items' representations disagree in directions that can be treated as high-frequency noise, so a sparse item-item semantic graph with a few rounds of Laplacian smoothing acts as a low-pass filter that suppresses those deviations. Interaction sequences, likewise, contain spurious clicks that show up as anomalous frequency components; transforming the item-embedding sequence with the Fast Fourier Transform, multiplying by a learnable c

What carries the argument

The load-bearing machinery is a pair of spectral filters plus an alignment loss. (1) Graph-based feature denoising: an item semantic graph built from cosine similarity with top-K neighbors and symmetrically normalized adjacency; repeated aggregation with a learnable adaptive weight implements Laplacian smoothing as a low-pass filter over feature noise, with the adaptive weight balancing self-features against neighbor aggregates. (2) Frequency-domain sequence denoising: the fused item-embedding sequence is mapped to the frequency domain via FFT, a learnable complex-valued filter modulates the spectrum element-wise, and IFFT returns a denoised sequence; a position-wise gate blends the filtered

Load-bearing premise

The frequency-domain sequence denoising module assumes that spurious interactions such as accidental clicks appear as high-frequency components in the FFT of the item-embedding sequence, and that a learnable spectral filter can separate them from genuine preferences; the paper offers no spectral analysis of real sequences, and since the FFT runs over integer positions rather than real interaction times, the frequency axis lacks direct temporal meaning.

What would settle it

A direct test would be to take real user sequences, inject known synthetic mis-clicks, compute the FFT spectrum of the item-embedding sequence, and check whether the injected noise concentrates in high-frequency bins and whether the learned filter specifically attenuates those bins; if the filter's attenuation does not track the injected noise, or if removing the frequency-domain module changes results only when injected noise is present, the core assumption would be falsified.

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

If this is right

  • If DDMSR is correct, feature-level denoising for recommendation does not require expensive generative models: structural low-pass filtering on a semantic graph suffices to improve pre-trained multi-modal representations.
  • Frequency-domain sequence filtering provides a single-shot, differentiable way to attenuate accidental clicks and other spurious interactions, integrated before a Transformer encoder.
  • Cross-modal contrastive alignment contributes to performance, but the paper's visualizations indicate graph denoising is the main driver of feature-distribution isotropy and cross-modal alignment.
  • The framework yields larger relative gains on rare and long-tail items compared with ID-based models, suggesting semantic neighbor aggregation imputes useful signals where interactions are scarce.
  • DDMSR trains roughly 2.5–3 times faster per epoch and uses 30–50 percent less GPU memory than a diffusion-based multi-modal denoising baseline while improving accuracy, supporting the claim that discriminative denoising is a practical alternative to generative denoising.

Where Pith is reading between the lines

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

  • The FFT-based sequence denoising operates on integer token positions rather than real interaction timestamps, so the 'frequency' axis measures positional periodicity, not temporal frequency; feeding real inter-click intervals as sample positions could test whether noise separation improves with true timing information.
  • If the high-frequency-noise assumption holds, the same frequency-domain filter could be applied to other behavioral sequences (search, browsing) or to user embeddings themselves, not just item sequences.
  • The observed isotropy after graph denoising suggests the module may reduce dimensional collapse in pre-trained embeddings; an implication is that the denoised features could make downstream recommenders more robust to embedding dimension choices.
  • The learnable adaptive weight in graph denoising resembles an attention mechanism over semantic neighborhoods; a possible extension is to replace the fixed top-K graph with a learned graph, although that would trade away the simplicity the paper emphasizes.

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

3 major / 5 minor

Summary. The paper proposes DDMSR, a multi-modal sequential recommendation framework with three components: (i) a graph-based feature denoising module (GFD) that applies Laplacian smoothing on an item-item semantic graph to suppress redundant pre-trained feature noise; (ii) a frequency-domain sequence denoising module (FSD) that applies FFT/IFFT with a learnable complex filter to attenuate spurious interactions; and (iii) a cross-modal contrastive alignment loss to harmonize textual and visual embeddings. Experiments on Beauty, Sports, Toys, and MicroLens claim consistent state-of-the-art improvements over ten baselines, with up to +19.33% Recall@20, plus ablation, hyperparameter sensitivity, efficiency, and long-tail analyses.

Significance. If the empirical claims withstand scrutiny, the work is valuable: it packages two lightweight denoising mechanisms—graph-based low-pass filtering and learnable spectral modulation—as an efficient alternative to diffusion-based denoising, with reported 2.5–3× training speedups and 30–50% memory reductions over DMMD4SR. The paper also provides code and a detailed appendix. However, the empirical foundation currently has reproducibility gaps that must be fixed before the SOTA claim can be accepted.

major comments (3)
  1. [Table 2 vs. Table 4 (Sec. 4.2 and A.3)] The same models and datasets show different scores in the two tables. For example, DDMSR Beauty R@20 is 0.1564 in Table 2 but 0.1579 in Table 4; DMMD4SR Beauty R@20 is 0.1171 vs 0.1129. Similar discrepancies appear for Sports, Toys, and MicroLens. Sec. 4.1.4 states that all final results are averaged over five independent runs, so these numbers should be identical unless the runs or settings differed, which is not described. This internal inconsistency directly undermines the reproducibility of the headline SOTA improvements. Please reconcile the tables and explain the source of the variation.
  2. [Sec. 4.1.4 and Table 2 (statistical reporting)] Table 2 marks many improvements with '*' and states statistical significance via paired t-tests (p<0.01), but no standard deviations, t-statistics, or details of the testing procedure are provided. Without variance measures, the reader cannot judge whether the reported gaps are meaningful, especially given the cross-table inconsistencies noted above. Report standard deviations for all metrics and describe the paired t-test setup (e.g., per-seed results, correction for multiple comparisons).
  3. [Sec. 3.3.2, Eq. (10)–(12)] The FSD module is motivated by the claim that accidental clicks appear as high-frequency components, but the FFT is taken over integer item positions, not real interaction timestamps, so the frequency axis has no direct temporal interpretation. The paper provides no spectral analysis of real sequences or any direct evidence that the learnable filter suppresses noise rather than performing an arbitrary data-dependent transformation. Since the filter is learned, the observed gains could arise from a generic learnable layer. Please add a diagnostic (e.g., visualize the learned filter, compare against fixed low-pass/high-pass filters, or run a synthetic noisy-sequence experiment) to substantiate the denoising mechanism.
minor comments (5)
  1. [Introduction (after 'Conference acronym XX')] The text contains an untranslated Chinese passage ('预训练模型...') and garbled figure captions. This appears to be a leftover from an early draft and should be cleaned up before submission.
  2. [Sec. 4.1.3 / 4.1.4 (baseline tuning)] The description of baseline hyperparameter tuning is vague ('following the search spaces specified in their respective original papers'). Please state explicitly whether the reported baseline numbers were produced locally with tuned hyperparameters on these datasets, or taken from prior publications. This is important for a fair comparison.
  3. [Eq. (13)] The gating network MLP in Eq. (13) is not specified; please give the hidden size and activation, or state that it is a single linear layer followed by sigmoid.
  4. [Figure 2 and Figure 5] The figures are dense and the text is small; consider enlarging the fonts and simplifying the diagrams for readability. Figure 5's KDE plots would benefit from clearer axis labels and a legend.
  5. [Sec. 3.2.2] The graph construction uses the same raw features that are later denoised. Please clarify whether the graph is built once and frozen, and whether the similarity matrix could introduce any circularity in the evaluation (it does not appear to, but this should be stated explicitly).

Circularity Check

0 steps flagged

No significant circularity: DDMSR's claimed gains are empirical evaluations on held-out data, and the sole self-citation is a non-load-bearing inspiration mention.

full rationale

The paper's central claim is an empirical comparison on held-out test data (Table 2) using leave-one-out evaluation over four public benchmarks. DDMSR's components are fully specified in Sections 3.2-3.4: graph construction from raw CLIP/RoBERTa features (Eq. 2-5), MoE adapters (Eq. 6-8), FFT/IFFT sequence filtering with a learned complex filter and gate (Eq. 10-14), Transformer encoder (Eq. 15-16), and the training objective (Eq. 17-19). None of these equations define a target quantity in terms of the quantity they are said to predict. The learnable filter W and gating weights are fit to the recommendation loss and then assessed on held-out interactions; ablations in Table 3 compare the full model against component-removed variants, so the contribution claims are not forced by construction. The only self-citation, [21], is cited alongside external works [29,48] as inspiration for frequency-domain denoising; it supplies no uniqueness theorem, no fitted parameter, and no ansatz on which the paper's equations depend. The 'noise is high-frequency' premise of the FSD module is an assumption, not a consequence of the model; if unsupported, that is a correctness risk, not circularity. The numeric discrepancies between Tables 2 and 4 affect reproducibility and verifiability, but they do not make the derivation equivalent to its inputs, so they do not count as circularity under the stated analysis criteria.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The method depends on several domain assumptions about where noise lives—in pre-trained features and in the frequency spectrum of sequences—that are plausible but not independently established. Standard math axioms cover the FFT formalism. No new physical or conceptual entities are introduced.

free parameters (5)
  • K (top-K semantic neighbors) = 5-20 per dataset (5 for Beauty/Toys, 10 for Sports, 20 for MicroLens)
    Searched in {5,10,15,20}; selected by validation performance per dataset, directly controls graph sparsity and denoising behavior.
  • L_f (number of FFT denoising layers) = 1-3 per dataset (2 for Beauty/MicroLens, 1 for Toys, 3 for Sports)
    Searched in {1,2,3}; optimal value varies by dataset, indicating the filter depth is tuned rather than derived.
  • lambda (contrastive loss weight) = 0.1
    Set to 0.1 based on preliminary experiments (§4.1.4); sensitivity analysis shows a sharp drop for larger values.
  • L_g (number of graph convolution layers) = 2
    Fixed by hand; no sensitivity analysis is reported for this choice.
  • temperature parameters tau_rec / tau_cl
    Used in Eq. (17)-(18) but their values are not reported; they are hand-chosen or defaulted and affect training behavior.
axioms (5)
  • domain assumption Noisy interactions appear as high-frequency components in the FFT of an item-embedding sequence.
    Invoked in §3.3.2; the paper gives intuition but no spectral analysis of real sequences to verify that mis-clicks are high-frequency rather than distributed across the spectrum.
  • domain assumption Laplacian smoothing over an item graph built from raw cosine similarities removes task-irrelevant feature noise.
    §3.2.2 assumes semantically similar items share coherent representations; the graph is constructed from the same noisy raw features, so aggregation could propagate noise unless the denoising assumption holds.
  • domain assumption Frozen CLIP/RoBERTa representations contain redundant, recommendation-irrelevant components that can be suppressed without losing predictive signal.
    This motivates the entire feature-denoising branch (§1, §3.2); it is plausible but not independently quantified.
  • standard math FFT/IFFT are exact invertible transforms with standard complexity.
    Appendix A.1 states standard discrete Fourier transform facts; these are not in dispute.
  • domain assumption The leave-one-out evaluation protocol and 5-core filtering produce a fair benchmark.
    §4.1.1 follows Hou et al. [10]; it is a conventional protocol but may not reflect deployment conditions such as true cold-start or online drift.

pith-pipeline@v1.3.0-alltime-deepseek · 22521 in / 16307 out tokens · 153565 ms · 2026-08-01T14:18:39.576257+00:00 · methodology

0 comments
read the original abstract

Multi-modal Sequential Recommendation (SR) incorporates rich side information (e.g., textual and visual features) to enhance dynamic user preference modeling. However, existing frameworks inevitably suffer from a \textbf{Dual-Noise Dilemma}: (1) \textit{Feature-level redundancy} stemming from the semantic gap between generic pre-trained representations and fine-grained recommendation intent; and (2) \textit{Sequence-level stochasticity} induced by spurious interactions such as accidental clicks. To break this bottleneck, we propose \textbf{DDMSR}, a novel \textbf{D}ual-level \textbf{D}enoising \textbf{M}ulti-modal \textbf{S}equential \textbf{R}ecommendation framework that systematically purifies signals from both feature-topological and sequence-frequency perspectives. Specifically, we first design a graph-based feature denoising module that leverages Laplacian smoothing on item semantic graphs as a structural low-pass filter, effectively suppressing high-frequency semantic noise while preserving salient features. For sequence purification, we introduce a frequency-domain sequence denoising module, utilizing the Fast Fourier Transform and a learnable frequency filter to adaptively modulate the interaction spectrum and attenuate anomalous signals. Furthermore, a multi-modal contrastive alignment objective is incorporated to bridge the heterogeneity gap and enforce cross-modal semantic consistency. Extensive experiments on four public benchmark datasets demonstrate that DDMSR consistently outperforms state-of-the-art baselines, providing a highly robust and efficient solution for multi-modal sequential recommendation. The source code is available at: ~\href{https://github.com/jluo00/DDMSR}{\textcolor{blue}{https://github.com/jluo00/DDMSR}}.

Figures

Figures reproduced from arXiv: 2607.18786 by Jie Luo, Qi Jin, Xinming Zhang.

Figure 1
Figure 1. Figure 1: Illustration of multi-level noise in multi-modal recommendation. (a) Feature-level noise: Pre-trained encoders (e.g., [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall framework of DDMSR. user 𝑢 ∈ U, their historical interaction sequence is chronologi￾cally ordered as 𝑆𝑢 = [𝑖1,𝑖2, . . . ,𝑖𝑛], where 𝑖𝑡 ∈ I denotes the 𝑡-th interacted item and 𝑛 = |𝑆𝑢 | is the sequence length. The goal of multi-modal sequential recommendation is to predict the next item 𝑖𝑛+1 that user 𝑢 is most likely to interact with, given 𝑆𝑢 and the associated multi-modal features: 𝑖𝑛+1 = arg ma… view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of textual (green) and visual (blue) [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effects of frequency-domain denoising layers. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Effects of the multi-modal fusion type. 0.17 0.15 0 N @) 0.13 -- RS u 0.11 cu a:: 0.09 [ -o- Beauty -o- Sports -o- Toys -o- Microlens J 0.1 A 0.5 1.0 0.070 0.062 0 N @) 0.054 ~ u 0.046 c z 0.038 0.1 A 0.5 1.0 [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Effects of the contrastive learning weight [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Performance improvement across different item [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

55 extracted references · 7 canonical work pages · 1 internal anchor

  1. [1]

    David H. Bailey. 1993. Computational Frameworks for the Fast Fourier Transform (Charles Van Loan).SIAM Rev.35, 1 (1993), 142–143. doi:10.1137/1035017

    work page doi:10.1137/1035017 1993

  2. [2]

    Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guil- laume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettle- moyer, and Veselin Stoyanov. 2020. Unsupervised Cross-lingual Representa- tion Learning at Scale. InProceedings of the 58th Annual Meeting of the Associa- tion for Computational Linguistics, ACL 2020, Online, Jul...

    doi:10.18653/v1/2020.acl-main.747 2020

  3. [3]

    Ziqiang Cui, Haolun Wu, Bowei He, Ji Cheng, and Chen Ma. 2024. Context Matters: Enhancing Sequential Recommendation with Context-aware Diffusion- based Contrastive Learning. InProceedings of the 33rd ACM International Con- ference on Information and Knowledge Management, CIKM 2024, Boise, ID, USA, October 21-25, 2024. ACM, 404–414. doi:10.1145/3627673.3679655

    arXiv 2024

  4. [4]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Associa- tion for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, ...

    doi:10.18653/v1/n19-1423 2019

  5. [5]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. In9th Interna- tional Conference on Learning Representations, IC...

    2021

  6. [6]

    Xinyu Du, Huanhuan Yuan, Pengpeng Zhao, Jianfeng Qu, Fuzhen Zhuang, Guanfeng Liu, Yanchi Liu, and Victor S. Sheng. 2023. Frequency Enhanced Hybrid Attention Network for Sequential Recommendation. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2023, Taipei, Taiwan, July 23-27, 2023. ...

    arXiv 2023

  7. [7]

    Junchen Fu, Xuri Ge, Xin Xin, Alexandros Karatzoglou, Ioannis Arapakis, Jie Wang, and Joemon M. Jose. 2024. IISAN: Efficiently Adapting Multimodal Repre- sentation for Sequential Recommendation with Decoupled PEFT. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2024, Washington DC, ...

    arXiv 2024

  8. [8]

    Xavier Glorot, Antoine Bordes, and Yoshua Bengio. 2011. Deep Sparse Rectifier Neural Networks. InProceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, AISTATS 2011, Fort Lauderdale, USA, April 11-13, 2011 (JMLR Proceedings). JMLR.org, 315–323. http://proceedings.mlr.press/ v15/glorot11a/glorot11a.pdf

    2011

  9. [9]

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk

  10. [10]

    Yupeng Hou, Shanlei Mu, Wayne Xin Zhao, Yaliang Li, Bolin Ding, and Ji-Rong Wen. 2022. Towards Universal Sequence Representation Learning for Recom- mender Systems. InKDD ’22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, August 14 - 18, 2022. ACM, 585–593. doi:10.1145/3534678.3539381

    arXiv 2022

  11. [11]

    Hengchang Hu, Wei Guo, Yong Liu, and Min-Yen Kan. 2023. Adaptive Multi- Modalities Fusion in Sequential Recommendation Systems. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management, CIKM 2023, Birmingham, United Kingdom, October 21-25, 2023. ACM, 843–853. doi:10.1145/3583780.3614775

    arXiv 2023

  12. [12]

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

  13. [13]

    Wang-Cheng Kang and Julian J. McAuley. 2018. Self-Attentive Sequential Rec- ommendation. InIEEE International Conference on Data Mining, ICDM 2018, Singapore, November 17-20, 2018. 197–206. doi:10.1109/ICDM.2018.00035

    arXiv 2018

  14. [14]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2015. Adam: A Method for Stochas- tic Optimization. In3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings. http://arxiv.org/abs/1412.6980

    Pith/arXiv arXiv 2015

  15. [15]

    Walid Krichene and Steffen Rendle. 2020. On Sampled Metrics for Item Rec- ommendation. InKDD ’20: The 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, CA, USA, August 23-27, 2020. 1748–

    2020

  16. [16]

    Yutong Li and Xinyi Zhang. 2025. MDSBR: Multimodal Denoising for Session- based Recommendation. InProceedings of the Nineteenth ACM Conference on Recommender Systems, RecSys 2025, Prague, Czech Republic, September 22-26, 2025. ACM, 268–278. https://doi.org/10.1145/3705328.3748061

    arXiv 2025

  17. [17]

    Zihao Li, Aixin Sun, and Chenliang Li. 2024. DiffuRec: A Diffusion Model for Sequential Recommendation.ACM Trans. Inf. Syst.42, 3 (2024), 66:1–66:28. https://doi.org/10.1145/3631116

    doi:10.1145/3631116 2024

  18. [18]

    Jiahao Liang, Xiangyu Zhao, Muyang Li, Zijian Zhang, Wanyu Wang, Haochen Liu, and Zitao Liu. 2023. MMMLP: Multi-modal Multilayer Perceptron for Se- quential Recommendations. InProceedings of the ACM Web Conference 2023, WWW 2023, Austin, TX, USA, 30 April 2023 - 4 May 2023. ACM, 1109–1117. doi:10.1145/3543507.3583378

    arXiv 2023

  19. [19]

    Qidong Liu, Jiaxi Hu, Yutian Xiao, Xiangyu Zhao, Jingtong Gao, Wanyu Wang, Qing Li, and Jiliang Tang. 2025. Multimodal Recommender Systems: A Survey. ACM Comput. Surv.57, 2 (2025), 26:1–26:17. doi:10.1145/3695461

    work page doi:10.1145/3695461 2025

  20. [20]

    Weihai Lu and Li Yin. 2025. DMMD4SR: Diffusion Model-based Multi-level Multimodal Denoising for Sequential Recommendation. InProceedings of the 33rd ACM International Conference on Multimedia, MM 2025, Dublin, Ireland, October 27-31, 2025. ACM, 6363–6372. https://doi.org/10.1145/3746027.3755861

    arXiv 2025

  21. [21]

    Jie Luo, Wenyu Zhang, Xinming Zhang, and Yuan Fang. 2026. Time-Aware Adaptive Side Information Fusion for Sequential Recommendation. InProceedings of the Nineteenth ACM International Conference on Web Search and Data Mining, WSDM 2026, Boise, ID, USA, February 22-26, 2026. ACM, 469–478. https://doi.org/ 10.1145/3773966.3778012

    arXiv 2026

  22. [22]

    Haokai Ma, Ruobing Xie, Lei Meng, Xin Chen, Xu Zhang, Leyu Lin, and Zhanhui Kang. 2024. Plug-In Diffusion Model for Sequential Recommenda- tion. InThirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, IAAI 2024, Fourteenth Symposium on Educational Advances in Art...

    work page doi:10.1609/aaai.v38i8.28736 2024

  23. [23]

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

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

  24. [24]

    Yongxin Ni, Yu Cheng, Xiangyan Liu, Junchen Fu, Youhua Li, Xiangnan He, Yongfeng Zhang, and Fajie Yuan. 2023. A Content-Driven Micro-Video Rec- ommendation Dataset at Scale.CoRRabs/2309.15379 (2023). arXiv:2309.15379 doi:10.48550/ARXIV.2309.15379

    Pith/arXiv arXiv doi:10.48550/arxiv.2309.15379 2023

  25. [25]

    Liwei Pan, Weike Pan, Meiyan Wei, Hongzhi Yin, and Zhong Ming. 2024. A Survey on Sequential Recommendation.CoRRabs/2412.12770 (2024). arXiv:2412.12770 doi:10.48550/ARXIV.2412.12770

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2412.12770 2024

  26. [26]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision. InProceedings of the 38th In- ternational Conference on Machine Learning, ICML 2021, 18-24 Jul...

    2021

  27. [27]

    Steffen Rendle, Christoph Freudenthaler, and Lars Schmidt-Thieme. 2010. Factor- izing personalized Markov chains for next-basket recommendation. InProceedings of the 19th International Conference on World Wide Web, WWW 2010, Raleigh, North Carolina, USA, April 26-30, 2010. 811–820. doi:10.1145/1772690.1772773

    arXiv 2010

  28. [28]

    Le, Geoffrey E

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc V. Le, Geoffrey E. Hinton, and Jeff Dean. 2017. Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer. In5th International Conference Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Jie Luo, Qi Jin, and Xinming Zhang on Learning Representations, ICLR 201...

    2017

  29. [29]

    Yehjin Shin, Jeongwhan Choi, Hyowon Wi, and Noseong Park. 2024. An Atten- tive Inductive Bias for Sequential Recommendation beyond the Self-Attention. In Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence, IAAI 2024, Four- teenth Symposium on Educational Advanc...

    work page doi:10.1609/aaai.v38i8 2024

  30. [30]

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang

  31. [31]

    Jiaxi Tang and Ke Wang. 2018. Personalized Top-N Sequential Recommendation via Convolutional Sequence Embedding. InProceedings of the Eleventh ACM International Conference on Web Search and Data Mining, WSDM 2018, Marina Del Rey, CA, USA, February 5-9, 2018. 565–573. doi:10.1145/3159652.3159656

    arXiv 2018

  32. [32]

    Aäron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation Learning with Contrastive Predictive Coding.CoRRabs/1807.03748 (2018). arXiv:1807.03748 http://arxiv.org/abs/1807.03748

    Pith/arXiv arXiv 2018

  33. [33]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All you Need. InAdvances in Neural Information Processing Systems 30: Annual Con- ference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA. 5998–6008. https://proceedings.neu...

    2017

  34. [34]

    Jinpeng Wang, Ziyun Zeng, Yunxiao Wang, Yuting Wang, Xingyu Lu, Tianxiang Li, Jun Yuan, Rui Zhang, Hai-Tao Zheng, and Shu-Tao Xia. 2023. MISSRec: Pre- training and Transferring Multi-modal Interest-aware Sequence Representation for Recommendation. InProceedings of the 31st ACM International Conference on Multimedia, MM 2023, Ottawa, ON, Canada, 29 October...

    arXiv 2023

  35. [35]

    Sheng, Mehmet A

    Shoujin Wang, Longbing Cao, Yan Wang, Quan Z. Sheng, Mehmet A. Orgun, and Defu Lian. 2022. A Survey on Session-based Recommender Systems.ACM Comput. Surv.54, 7 (2022), 154:1–154:38. doi:10.1145/3465401

    work page doi:10.1145/3465401 2022

  36. [36]

    Sheng, and Mehmet A

    Shoujin Wang, Liang Hu, Yan Wang, Longbing Cao, Quan Z. Sheng, and Mehmet A. Orgun. 2019. Sequential Recommender Systems: Challenges, Progress and Prospects. InProceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI 2019, Macao, China, August 10-16, 2019. 6332–6338. doi:10.24963/IJCAI.2019/883

    work page doi:10.24963/ijcai.2019/883 2019

  37. [37]

    Tongzhou Wu, Yuhao Wang, Maolin Wang, Chi Zhang, and Xiangyu Zhao. 2025. Empowering Denoising Sequential Recommendation with Large Language Model Embeddings. InProceedings of the 34th ACM International Conference on Informa- tion and Knowledge Management, CIKM 2025, Seoul, Republic of Korea, November 10-14, 2025. ACM, 3427–3437. https://doi.org/10.1145/37...

    arXiv 2025

  38. [38]

    Lanling Xu, Zhen Tian, Bingqian Li, Junjie Zhang, Daoyuan Wang, Hongyu Wang, Jinpeng Wang, Sheng Chen, and Wayne Xin Zhao. 2024. Sequence-level Semantic Representation Fusion for Recommender Systems. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management, CIKM 2024, Boise, ID, USA, October 21-25, 2024. ACM, 5015–502...

    arXiv 2024

  39. [39]

    Guipeng Xv, Xinyu Li, Ruobing Xie, Chen Lin, Chong Liu, Feng Xia, Zhanhui Kang, and Leyu Lin. 2024. Improving Multi-modal Recommender Systems by Denoising and Aligning Multi-modal Content and User Feedback. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2024, Barcelona, Spain, August 25-29, 2024. ACM, 3645–3656...

    arXiv 2024

  40. [40]

    Jose, and Xiangnan He

    Fajie Yuan, Alexandros Karatzoglou, Ioannis Arapakis, Joemon M. Jose, and Xiangnan He. 2019. A Simple Convolutional Generative Network for Next Item Recommendation. InProceedings of the Twelfth ACM International Conference on Web Search and Data Mining, WSDM 2019, Melbourne, VIC, Australia, February 11-15, 2019. 582–590. doi:10.1145/3289600.3290975

    arXiv 2019

  41. [41]

    Zheng Yuan, Fajie Yuan, Yu Song, Youhua Li, Junchen Fu, Fei Yang, Yunzhu Pan, and Yongxin Ni. 2023. Where to Go Next for Recommender Systems? ID- vs. Modality-based Recommender Models Revisited. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2023, Taipei, Taiwan, July 23-27, 2023. 2...

    arXiv 2023

  42. [42]

    Theory and Application of Digital Signal Processing

    C. K. Yuen. 1978. Review of "Theory and Application of Digital Signal Processing" by Lawrence R. Rabiner and Bernard Gold.IEEE Trans. Syst. Man Cybern.8, 2 (1978), 146. doi:10.1109/TSMC.1978.4309918

    arXiv 1978

  43. [43]

    Chi Zhang, Yantong Du, Xiangyu Zhao, Qilong Han, Rui Chen, and Li Li. 2022. Hierarchical Item Inconsistency Signal Learning for Sequence Denoising in Se- quential Recommendation. InProceedings of the 31st ACM International Conference on Information & Knowledge Management, Atlanta, GA, USA, October 17-21, 2022. ACM, 2508–2518. https://doi.org/10.1145/35118...

    arXiv 2022

  44. [44]

    Chi Zhang, Qilong Han, Rui Chen, Xiangyu Zhao, Peng Tang, and Hong- tao Song. 2024. SSDRec: Self-Augmented Sequence Denoising for Sequential Recommendation. In40th IEEE International Conference on Data Engineering, ICDE 2024, Utrecht, The Netherlands, May 13-16, 2024. IEEE, 803–815. https: //doi.org/10.1109/ICDE60146.2024.00067

    arXiv 2024

  45. [45]

    Lingzi Zhang, Xin Zhou, Zhiwei Zeng, and Zhiqi Shen. 2025. Multimodal Pre-training for Sequential Recommendation via Contrastive Learning.Trans. Recomm. Syst.3, 1 (2025), 1–23. https://doi.org/10.1145/3682075

    doi:10.1145/3682075 2025

  46. [46]

    Shengzhe Zhang, Liyi Chen, Dazhong Shen, Chao Wang, and Hui Xiong. 2025. Hierarchical Time-Aware Mixture of Experts for Multi-Modal Sequential Rec- ommendation. InProceedings of the ACM on Web Conference 2025, WWW 2025, Sydney, NSW, Australia, 28 April 2025- 2 May 2025. ACM, 3672–3682. doi:10.1145/3696410.3714676

    arXiv 2025

  47. [47]

    Wayne Xin Zhao, Shanlei Mu, Yupeng Hou, Zihan Lin, Yushuo Chen, Xingyu Pan, Kaiyuan Li, Yujie Lu, Hui Wang, Changxin Tian, Yingqian Min, Zhichao Feng, Xinyan Fan, Xu Chen, Pengfei Wang, Wendi Ji, Yaliang Li, Xiaoling Wang, and Ji- Rong Wen. 2021. RecBole: Towards a Unified, Comprehensive and Efficient Frame- work for Recommendation Algorithms. InCIKM ’21:...

    arXiv 2021

  48. [48]

    Kun Zhou, Hui Yu, Wayne Xin Zhao, and Ji-Rong Wen. 2022. Filter-enhanced MLP is All You Need for Sequential Recommendation. InWWW ’22: The ACM Web Conference 2022, Virtual Event, Lyon, France, April 25 - 29, 2022. 2388–2399. doi:10.1145/3485447.3512111

    arXiv 2022

  49. [49]

    Tian Zhou, Ziqing Ma, Xue Wang, Qingsong Wen, Liang Sun, Tao Yao, Wotao Yin, and Rong Jin. 2022. FiLM: Frequency improved Legendre Mem- ory Model for Long-term Time Series Forecasting. InAdvances in Neural Information Processing Systems 35: Annual Conference on Neural Informa- tion Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, Novem- ber 28...

    2022

  50. [50]

    Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. 2022. FEDformer: Frequency Enhanced Decomposed Transformer for Long-term Series Forecasting. InInternational Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA (Proceedings of Machine Learning Research). PMLR, 27268–27286. https://proceedings.mlr.press...

    2022

  51. [1757]

    doi:10.1145/3394486.3403226

    arXiv

  52. [2015]

    InProceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval, Santiago, Chile, August 9-13, 2015

    Image-Based Recommendations on Styles and Substitutes. InProceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval, Santiago, Chile, August 9-13, 2015. 43–52. doi:10.1145/ 2766462.2767755

    arXiv 2015

  53. [2016]

    In4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings

    Session-based Recommendations with Recurrent Neural Networks. In4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings. http://arxiv.org/abs/1511.06939

    Pith/arXiv arXiv 2016

  54. [2019]

    InProceedings of the 28th ACM International Con- ference on Information and Knowledge Management, CIKM 2019, Beijing, China, November 3-7, 2019

    BERT4Rec: Sequential Recommendation with Bidirectional Encoder Repre- sentations from Transformer. InProceedings of the 28th ACM International Con- ference on Information and Knowledge Management, CIKM 2019, Beijing, China, November 3-7, 2019. 1441–1450. doi:10.1145/3357384.3357895

    arXiv 2019

  55. [2024]

    InProceedings of the 32nd ACM International Conference on Multimedia, MM 2024, Melbourne, VIC, Australia, 28 October 2024 - 1 November 2024

    DiffMM: Multi-Modal Diffusion Model for Recommendation. InProceedings of the 32nd ACM International Conference on Multimedia, MM 2024, Melbourne, VIC, Australia, 28 October 2024 - 1 November 2024. ACM, 7591–7599. https: //doi.org/10.1145/3664647.3681498

    arXiv 2024