Pith. sign in

REVIEW 4 major objections 5 minor 39 references

EGRA:Toward Enhanced Behavior Graphs and Representation Alignment for Multimodal Recommendation

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims that EGRA outperforms state-of-the-art multimodal recommendation methods on five benchmarks by building the item-item graph from pretrained embeddings instead of raw modality features, and by adjusting modality-behavior ali

desk verdict Plausible graph-enhancement idea, but the SOTA claim hinges on an undisclosed pretraining split that could be test-label leakage. read the letter →

arxiv 2508.16170 v1 pith:NROALDQF submitted 2025-08-22 cs.IR cs.AI

Xiaoxiong Zhang , Xin Zhou , Zhiwei Zeng , Yongjie Wang , Dusit Niyato , Zhiqi Shen This is my paper
classification cs.IRcs.AI
keywords multimodalrecommendationbehaviorgraphenhancementitem-itemsemanticpretrainedembeddingsdynamicalignmentweightinglong-tailconvolution
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

EGRA is a multimodal recommendation method that tries to fix two things: the way item-item links are added to the user-item graph, and the way modality and behavior representations are pulled together. Instead of linking items by raw visual or textual feature similarity, it builds item-item edges from embeddings produced by a pretrained multimodal recommender, which already mix collaborative and modality signals, and it feeds those edges into the behavior graph before graph convolution. It then aligns modality and behavior embeddings with a contrastive loss whose weight is different for every entity and grows over training epochs. On Baby, Sports, Clothing, MicroLens and Elec, the paper reports consistent top-K Recall and NDCG improvements over ten baselines, with relative gains over the strongest baseline reaching 10.65%. A careful reader should care because the two components are separable and each contributes in ablations, suggesting the gains are not from a single trick.

What carries the argument

Two mechanisms carry the argument. The Enhanced Behavior Graph is built from the top-H cosine neighbors in the pretrained MGCN item embeddings, converted to binary edges and symmetrized into the user-item adjacency matrix before LightGCN-style convolution; this injects collaborative-plus-modality semantics directly into graph propagation. The Bi-Level Dynamic Alignment Weighting sets the alignment loss weight as the product of an epoch-wise ramp (lambda_min increasing to lambda_max over P warm-up epochs) and an entity-wise softmax over alignment discrepancy, so poorly aligned users and items receive stronger alignment pressure early against unstable representations. Together they make both t

What would settle it

Measure whether top-H neighbors from the pretrained embeddings are more predictive of held-out co-interaction than random item edges with the same degree distribution. If a variant of EGRA with random item-item edges of identical degree matches EGRA's Recall@20 on Baby or Sports, the semantic content of the pretrained graph is not the source of the gain. A second check: pretrain MGCN using only the training split's interactions and rerun EGRA; if most of the improvement disappears, the pretrained model's exposure to evaluation interactions is implicated.

Watch

Extended reading notes

Core claim

The paper's central claim is that multimodal recommendation accuracy improves when the item-item graph used to fight behavior-graph sparsity is constructed from the learned embedding space of a pretrained multimodal recommender rather than from raw modality features. The pretrained model chosen is MGCN, whose embeddings already contain both collaborative and modality-aware information; EGRA takes the top-5 cosine neighbors for each item as binary edges, symmetrizes them with the interaction matrix, and runs LightGCN over the combined graph. On the representation side, EGRA claims that fixed, uniform alignment strength leaves many entities misaligned; it therefore weights the contrastive alig

Load-bearing premise

The pretrained MGCN embeddings' top-K neighbors are genuine item similarities that help the downstream graph, not popularity artifacts or memorized interactions from the same data used to evaluate.

Editorial extensions

If this is right

  • If the pretrained-embedding graph is the cause of the gains, any MMR method that already produces fused collaborative-modality embeddings can serve as a graph constructor, not only as a feature extractor.
  • The long-tail experiments imply graph enhancement is disproportionately helpful for items with fewer interactions, since tail groups show larger gains over the no-enhancement variant.
  • Because replacing MGCN with a stronger pretrained model changes results only marginally, the method's performance appears to saturate in the graph-construction step rather than in the pretrained model's embedding quality.
  • The reported per-epoch training time is lower than GUME's, and total overhead including pretraining stays below 9% on the two largest datasets, so the extra graph is cheap enough for routine training.

Reading between the lines

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

  • Editorial inference: a direct control experiment with random same-degree item-item edges would test whether the semantic content of the pretrained embeddings is load-bearing or whether any graph densification helps; the paper does not run this control.
  • Editorial inference: because the pretrained MGCN is fit on the same interaction data used to train and evaluate EGRA, a split-aware pretraining protocol could separate genuine collaborative signal transfer from leakage; the paper does not address this.
  • Editorial inference: the entity-wise weighting rule, which assigns stronger alignment to more misaligned entities, resembles an adaptive curriculum and could transfer to other contrastive alignment tasks outside recommendation, such as multimodal retrieval.
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

4 major / 5 minor

Summary. The paper proposes EGRA, a multimodal recommendation framework with two main contributions: (1) an enhanced behavior graph that augments the user–item interaction graph with an item–item graph built from the item embeddings of a pretrained MGCN model (Section IV-A), and (2) a bi-level dynamic alignment weighting mechanism that adjusts modality–behavior alignment strength per entity and per training epoch (Section IV-C2). The method is evaluated on Baby, Sports, Clothing, MicroLens, and Elec against ten baselines, reporting consistent improvements over GUME with relative gains up to 10.65% (Table I). Additional experiments cover long-tail performance, ablations of the two components, a comparison with GUME's graph-enhancement strategy, sensitivity to the pretrained backbone, training efficiency, and hyperparameter analysis.

Significance. If the empirical claims hold, the paper makes a useful practical contribution: replacing raw-modality item–item graphs with graphs derived from a pretrained multimodal recommender is simple, transferable to several backbones (Table III), and brings consistent gains; the dynamic alignment weighting is also well motivated. The method is specified in enough detail to reimplement, and the ablations are internally consistent: removing the enhanced behavior graph or the alignment weighting degrades performance in the expected direction. The efficiency analysis is transparent about the extra pretraining cost. However, the central empirical claim is currently not convincing because the pretraining protocol for the MGCN backbone is not stated, and all reported numbers come from single runs without variance or significance testing.

major comments (4)
  1. [§IV-A1, §V-A] The pretraining split for MGCN is never stated. The item–item graph S^pt is computed from E^pt, which is obtained by pretraining MGCN on user–item interactions. If MGCN is pretrained on the full interaction matrix before the 8:1:1 split, then for a test interaction (u,i), item i can be connected in G_e to other items that co-occur with u in the training portion because the pretraining model already saw those labels. This would give EGRA access to test-informed graph structure that baselines such as GUME, which build item–item edges from raw modality features, do not have, and would directly explain the large gains on MicroLens and Elec in Table I. The authors must state the exact pretraining split and, ideally, provide an experiment where S^pt is built from an MGCN trained only on the 80% training split, together with an edge-quality analysis (e.g., precision of top-H neighbors against h
  2. [Table I, §V-B] All performance numbers are single runs without standard deviations or significance tests. The relative gains over GUME in Table I range from 1.83% to 10.65%, but with no variance information it is impossible to determine whether the claimed 'significant outperformance' is beyond run-to-run noise. For a claim of state-of-the-art results, the authors should report mean ± std over at least three seeds and perform paired significance tests for the key comparisons.
  3. [§V-A] The hyperparameter values selected by grid search are not reported. The text lists the search ranges for λ_min, λ_max, P, τ1, and τ2/3, and states that H and the number of semantic graph layers are fixed per dataset, but the actual configurations used to produce Table I are absent. Without this information the experiments are not reproducible, and the risk of selection bias in the SOTA comparisons cannot be assessed. Please include a table of the chosen hyperparameter values for each dataset.
  4. [§V-C, §V-D] The evidence for the two main components is incomplete. The ablation study in Table II is reported on only three of the five datasets, and the comparison of graph-enhancement strategies in Table III reports only R@20. Moreover, Figure 2, which is central to the long-tail claim, appears in the manuscript as unreadable glyph/codepoint artifacts rather than as plots. To support the 'consistent' and 'generalization' claims, all metrics should be reported for all datasets, with significance tests, and the long-tail figure should be legible.
minor comments (5)
  1. [Eq. (14)] The entity-wise softmax weights are normalized within each mini-batch, so the resulting alignment weights depend on the batch composition. The batch size is not reported; this should be stated and its sensitivity discussed.
  2. [References] Reference [39] is a duplicate of [12] (the GUME paper). Please unify the citations.
  3. [Eq. (2)] The retained item–item edges in S^pt are all assigned weight 1, discarding the cosine similarity magnitude. Since S^pt is claimed to carry both collaborative and modality-aware semantics, the binarization step deserves a brief justification or an ablation.
  4. [§V-F] The efficiency comparison adds MGCN pretraining time to EGRA's total but does not account for any analogous pretraining or tuning cost for baselines such as GUME. The comparison would be fairer if the total wall-clock time to reach the reported result were reported for every method.
  5. [§V-A] The sentence 'we adopt a pre-training strategy [28]' cites GraphDA, but the paper does not explain whether GraphDA's pretraining-split protocol is used unchanged. Please clarify the adaptation explicitly.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation found; the central SOTA claim rests on external baselines and ablations, though the MGCN pretraining split is unspecified.

full rationale

The claimed derivation is not circular. EGRA's enhanced graph is constructed as S^pt_{i,j}=Cos(E_{i,pt}, E_{j,pt}) from pretrained MGCN embeddings (Eq. 1), binarized to Top-H edges (Eq. 2), and merged with the interaction graph R into G_e; the behavior encoder then runs LightGCN on G_e (Eqs. 3-4), and the final score is an inner product of learned representations (Eq. 22). No equation defines the output in terms of itself, and no fitted parameter is relabeled as a prediction: the dynamic alignment weights (Eqs. 14-16) modulate a contrastive loss and are not test-set targets. The empirical SOTA claim is supported by comparisons to ten external baselines (Table I), component ablations (Table II), backbone-transfer tests (Table III), and pretraining-backbone substitution (Table IV), so it is not forced by a self-citation chain. The paper does contain self-citations ([5], [6], [34], [40]), but they appear as background, Top-K implementation, and framework references; none carries the load-bearing argument, which instead cites external GraphDA [28] for the pretraining strategy and external MGCN [11] as backbone. The one legitimate concern, flagged in Section V-A, is that the MGCN pretraining split is never stated: the protocol only describes the 8:1:1 split for 'the general evaluation.' If pretraining used test interactions, S^pt could leak test labels; this is a correctness/leakage risk that the paper should address, but it is not an in-text circular reduction and cannot be demonstrated from the manuscript. Overall circularity score 2 reflects the minor self-citations and the unspecified split, not a circular derivation.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The method rests on several domain assumptions about the reliability of pretrained embeddings and graph-based propagation, plus a set of grid-searched hyperparameters whose chosen values are not reported. No new physical or conceptual entities are introduced.

free parameters (7)
  • lambda_min = 0, 0.005, or 0.01 (per dataset, not reported)
    Initial epoch-wise alignment weight; grid searched in Section V-A and analyzed in Table VI.
  • lambda_max = 0.02, 0.03, or 0.04 (not reported per dataset)
    Upper bound of epoch-wise alignment weight; grid searched in Section V-A.
  • P = 5, 10, 15, or 20 (not reported per dataset)
    Number of warm-up epochs after which the epoch-wise weight is fixed; grid searched in Section V-A.
  • tau1 = 0.6 to 2.0 (not reported per dataset)
    Temperature for entity-wise softmax weighting in Eq. (14); grid searched in Section V-A.
  • tau2/tau3 = 0.1, 0.2, or 0.3 (not reported per dataset)
    Temperatures for the contrastive alignment loss in Eq. (17); grid searched in Section V-A.
  • H = 5 (set for all datasets, analyzed over 1-9)
    Number of top-K neighbors for the enhanced item-item graph; chosen via parameter analysis in Figure 3.
  • Semantic graph layers = 2 for Baby, 1 for others
    Number of graph convolution layers on the item-item semantic graph, selected per dataset in Section V-A.
assumptions (5)
  • domain assumption Cosine similarity in pretrained MGCN embedding space is a valid measure of item relevance for graph construction.
    Section IV-A1, Eqs. (1)-(2): item-item edges are built from top-K cosine neighbors of E_pt. If pretrained embeddings do not reflect semantic or collaborative relevance, the enhanced graph injects false edges.
  • domain assumption The pretrained MGCN is trained on the same interaction data in a way that does not leak test information.
    A pre-training strategy from [28] is adopted, but the paper does not explicitly state the split used for pretraining. Standard 8:1:1 protocol is implied.
  • domain assumption LightGCN propagation on the enhanced behavior graph preserves useful signal despite binary edge weights.
    The enhanced graph is encoded with normalized LightGCN (Eq. (3)); binary weights remove similarity magnitude, which may lose information.
  • domain assumption The disentanglement-based weighted fusion (Eqs. (11)-(13)) separates shared and modality-exclusive semantics as intended.
    The fusion subtracts a shared representation from modality features to get exclusive features; the validity of this decomposition is assumed.
  • domain assumption The interaction-aware alignment loss (Eq. (17)) does not distort the user-item interaction structure.
    The loss pulls modality and behavior representations together while using interacted items as anchors; the paper assumes this preserves interaction geometry.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EGRA:Toward Enhanced Behavior Graphs and Representation Alignment for Multimodal Recommendation." pith.science (2026). https://pith.science/paper/NROALDQF

@misc{pith2026250816170,
  author       = {Pith},
  title        = {Pith review of: EGRA:Toward Enhanced Behavior Graphs and Representation Alignment for Multimodal Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NROALDQF}},
  note         = {Machine review of arXiv:2508.16170}
}
read the original abstract

MultiModal Recommendation (MMR) systems have emerged as a promising solution for improving recommendation quality by leveraging rich item-side modality information, prompting a surge of diverse methods. Despite these advances, existing methods still face two critical limitations. First, they use raw modality features to construct item-item links for enriching the behavior graph, while giving limited attention to balancing collaborative and modality-aware semantics or mitigating modality noise in the process. Second, they use a uniform alignment weight across all entities and also maintain a fixed alignment strength throughout training, limiting the effectiveness of modality-behavior alignment. To address these challenges, we propose EGRA. First, instead of relying on raw modality features, it alleviates sparsity by incorporating into the behavior graph an item-item graph built from representations generated by a pretrained MMR model. This enables the graph to capture both collaborative patterns and modality aware similarities with enhanced robustness against modality noise. Moreover, it introduces a novel bi-level dynamic alignment weighting mechanism to improve modality-behavior representation alignment, which dynamically assigns alignment strength across entities according to their alignment degree, while gradually increasing the overall alignment intensity throughout training. Extensive experiments on five datasets show that EGRA significantly outperforms recent methods, confirming its effectiveness.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 33 canonical work pages

  1. [6]

    Semantic item graph enhancement for multimodal recommendation,

    X. Zhang, X. Zhou, Z. Zeng, D. Niyato, and Z. Shen, “Semantic item graph enhancement for multimodal recommendation,”arXiv preprint arXiv:2508.06154, 2025

    arXiv 2025

  2. [1]

    Multimodal hierarchical graph collaborative filtering for multimedia-based recommendation,

    K. Liu, F. Xue, S. Li, S. Sang, and R. Hong, “Multimodal hierarchical graph collaborative filtering for multimedia-based recommendation,” IEEE Transactions on Computational Social Systems, vol. 11, no. 1, pp. 216–227, 2022

    work page 2022

  3. [2]

    Multi-modal discrete collaborative filtering for efficient cold-start recommendation,

    Y . Xu, L. Zhu, Z. Cheng, J. Li, Z. Zhang, and H. Zhang, “Multi-modal discrete collaborative filtering for efficient cold-start recommendation,” IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 1, pp. 741–755, 2021

    work page 2021

  4. [3]

    Music recommendation by unified hypergraph: combining social media information and music content,

    J. Bu, S. Tan, C. Chen, C. Wang, H. Wu, L. Zhang, and X. He, “Music recommendation by unified hypergraph: combining social media information and music content,” inProceedings of the 18th ACM international conference on Multimedia, 2010, pp. 391–400

    work page 2010

  5. [4]

    Multimodal machine learning: A survey and taxonomy,

    T. Baltru ˇsaitis, C. Ahuja, and L.-P. Morency, “Multimodal machine learning: A survey and taxonomy,”IEEE transactions on pattern anal- ysis and machine intelligence, vol. 41, no. 2, pp. 423–443, 2018

    2018

  6. [5]

    Learning item representa- tions directly from multimodal features for effective recommendation,

    X. Zhou, X. Zhang, D. Niyato, and Z. Shen, “Learning item representa- tions directly from multimodal features for effective recommendation,” arXiv preprint arXiv:2505.04960, 2025

    arXiv 2025

  7. [7]

    Mining latent structures for multimedia recommendation,

    J. Zhang, Y . Zhu, Q. Liu, S. Wu, S. Wang, and L. Wang, “Mining latent structures for multimedia recommendation,” inProceedings of the 29th ACM International Conference on Multimedia, ser. MM ’21. New York, NY , USA: Association for Computing Machinery, 2021, p. 3872–3880

    work page 2021

  8. [8]

    A tale of two graphs: Freezing and denoising graph structures for multimodal recommendation,

    X. Zhou and Z. Shen, “A tale of two graphs: Freezing and denoising graph structures for multimodal recommendation,” inProceedings of the 31st ACM International Conference on Multimedia, ser. MM ’23. Association for Computing Machinery, 2023, p. 935–943

    work page 2023

Show all 39 references
  1. [9]

    Latent structure mining with contrastive modality fusion for multimedia rec- ommendation,

    J. Zhang, Y . Zhu, Q. Liu, M. Zhang, S. Wu, and L. Wang, “Latent structure mining with contrastive modality fusion for multimedia rec- ommendation,”IEEE Trans. on Knowl. and Data Eng., vol. 35, no. 9, p. 9154–9167, Sep. 2023

    2023

  2. [10]

    Improving multi-modal recommender systems by denoising and align- ing multi-modal content and user feedback,

    G. Xv, X. Li, R. Xie, C. Lin, C. Liu, F. Xia, Z. Kang, and L. Lin, “Improving multi-modal recommender systems by denoising and align- ing multi-modal content and user feedback,” inProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 3645–3656

    2024

  3. [11]

    Multi-view graph convolutional network for multimedia recommendation,

    P. Yu, Z. Tan, G. Lu, and B.-K. Bao, “Multi-view graph convolutional network for multimedia recommendation,” inProceedings of the 31st ACM International Conference on Multimedia, ser. MM ’23. New York, NY , USA: Association for Computing Machinery, 2023, p. 6576–6585

    2023

  4. [13]

    Dualgnn: Dual graph neural network for multimedia recommendation,

    Q. Wang, Y . Wei, J. Yin, J. Wu, X. Song, and L. Nie, “Dualgnn: Dual graph neural network for multimedia recommendation,”Trans. Multi., vol. 25, p. 1074–1084, Jan. 2023

    2023

  5. [14]

    Vbpr: visual bayesian personalized ranking from implicit feedback,

    R. He and J. McAuley, “Vbpr: visual bayesian personalized ranking from implicit feedback,” inProceedings of the Thirtieth AAAI Conference on Artificial Intelligence, ser. AAAI’16. AAAI Press, 2016, p. 144–150

    2016

  6. [15]

    Visually-aware fashion recommendation and design with generative image models,

    W.-C. Kang, C. Fang, Z. Wang, and J. McAuley, “Visually-aware fashion recommendation and design with generative image models,”2017 IEEE International Conference on Data Mining (ICDM), pp. 207–216, 2017

    2017

  7. [16]

    Deepstyle: Learning user preferences for visual recommendation,

    Q. Liu, S. Wu, and L. Wang, “Deepstyle: Learning user preferences for visual recommendation,” inProceedings of the 40th international acm sigir conference on research and development in information retrieval, 2017, pp. 841–844

    2017

  8. [17]

    Personalized fashion recommendation with visual explanations based on multimodal attention network: Towards visually explainable rec- ommendation,

    X. Chen, H. Chen, H. Xu, Y . Zhang, Y . Cao, Z. Qin, and H. Zha, “Personalized fashion recommendation with visual explanations based on multimodal attention network: Towards visually explainable rec- ommendation,” inProceedings of the 42nd International ACM SIGIR Conference on...

    2019

  9. [18]

    Mmgcn: Multi-modal graph convolution network for personalized recommenda- tion of micro-video,

    Y . Wei, X. Wang, L. Nie, X. He, R. Hong, and T.-S. Chua, “Mmgcn: Multi-modal graph convolution network for personalized recommenda- tion of micro-video,” inProceedings of the 27th ACM International Conference on Multimedia, ser. MM ’19. Association for Computing Machinery, 20...

    2019

  10. [19]

    Graph-refined convolutional network for multimedia recommendation with implicit feedback,

    Y . Wei, X. Wang, L. Nie, X. He, and T.-S. Chua, “Graph-refined convolutional network for multimedia recommendation with implicit feedback,” 10 2020, pp. 3541–3549

    2020

  11. [20]

    Bootstrap latent representations for multi-modal recommen- dation,

    X. Zhou, H. Zhou, Y . Liu, Z. Zeng, C. Miao, P. Wang, Y . You, and F. Jiang, “Bootstrap latent representations for multi-modal recommen- dation,” inProceedings of the ACM Web Conference 2023, ser. WWW ’23. Association for Computing Machinery, 2023, p. 845–854

    2023

  12. [21]

    Lgmrec: Local and global graph learning for multimodal recommendation,

    Z. Guo, J. Li, L. Guohui, C. Wang, S. Shi, and B. Ruan, “Lgmrec: Local and global graph learning for multimodal recommendation,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, pp. 8454– 8462, 03 2024

    2024

  13. [22]

    Challenging the long tail recommendation,

    H. Yin, B. Cui, J. Li, J. Yao, and C. Chen, “Challenging the long tail recommendation,”Proceedings of the VLDB Endowment, vol. 5, no. 9, 2012

    2012

  14. [23]

    Dropoutnet: Addressing cold start in recommender systems,

    M. V olkovs, G. Yu, and T. Poutanen, “Dropoutnet: Addressing cold start in recommender systems,”Advances in neural information processing systems, vol. 30, 2017

    2017

  15. [24]

    Melu: Meta-learned user preference estimator for cold-start recommendation,

    H. Lee, J. Im, S. Jang, H. Cho, and S. Chung, “Melu: Meta-learned user preference estimator for cold-start recommendation,” inProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 2019, pp. 1073–1082

    2019

  16. [25]

    Long-tail session-based recommendation,

    S. Liu and Y . Zheng, “Long-tail session-based recommendation,” in Proceedings of the 14th ACM conference on recommender systems, 2020, pp. 509–514

    2020

  17. [26]

    A model of two tales: Dual transfer learning framework for improved long- tail item recommendation,

    Y . Zhang, D. Z. Cheng, T. Yao, X. Yi, L. Hong, and E. H. Chi, “A model of two tales: Dual transfer learning framework for improved long- tail item recommendation,” inProceedings of the web conference 2021, 2021, pp. 2220–2231

    2021

  18. [27]

    Improving long-tail item recom- mendation with graph augmentation,

    S. Luo, C. Ma, Y . Xiao, and L. Song, “Improving long-tail item recom- mendation with graph augmentation,” inProceedings of the 32nd ACM international conference on information and knowledge management, 2023, pp. 1707–1716

    2023

  19. [28]

    Graph collaborative signals denoising and augmentation for recommendation,

    Z. Fan, K. Xu, Z. Dong, H. Peng, J. Zhang, and P. S. Yu, “Graph collaborative signals denoising and augmentation for recommendation,” inProceedings of the 46th international ACM SIGIR conference on research and development in information retrieval, 2023, pp. 2037–2041

    2023

  20. [29]

    Diffusion augmentation for sequential recommendation,

    Q. Liu, F. Yan, X. Zhao, Z. Du, H. Guo, R. Tang, and F. Tian, “Diffusion augmentation for sequential recommendation,” inProceedings of the 32nd ACM International conference on information and knowledge management, 2023, pp. 1576–1586

    2023

  21. [30]

    Linrec: Linear attention mechanism for long- term sequential recommender systems,

    L. Liu, L. Cai, C. Zhang, X. Zhao, J. Gao, W. Wang, Y . Lv, W. Fan, Y . Wang, M. Heet al., “Linrec: Linear attention mechanism for long- term sequential recommender systems,” inProceedings of the 46th International ACM SIGIR Conference on Research and Development in Informatio...

    2023

  22. [31]

    Coral: Collaborative retrieval-augmented large language models im- prove long-tail recommendation,

    J. Wu, C.-C. Chang, T. Yu, Z. He, J. Wang, Y . Hou, and J. McAuley, “Coral: Collaborative retrieval-augmented large language models im- prove long-tail recommendation,” inProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 3391–3401

    2024

  23. [32]

    Llm-esr: Large language models enhancement for long-tailed sequential recommendation,

    Q. Liu, X. Wu, Y . Wang, Z. Zhang, F. Tian, Y . Zheng, and X. Zhao, “Llm-esr: Large language models enhancement for long-tailed sequential recommendation,”Advances in Neural Information Processing Systems, vol. 37, pp. 26 701–26 727, 2024

    2024

  24. [33]

    Integrating large language models into recommendation via mutual augmentation and adaptive aggregation,

    S. Luo, Y . Yao, B. He, Y . Huang, A. Zhou, X. Zhang, Y . Xiao, M. Zhan, and L. Song, “Integrating large language models into recommendation via mutual augmentation and adaptive aggregation,”arXiv preprint arXiv:2401.13870, 2024

    2024

  25. [34]

    Communication- efficient federated knowledge graph embedding with entity-wise top-k sparsification,

    X. Zhang, Z. Zeng, X. Zhou, D. Niyato, and Z. Shen, “Communication- efficient federated knowledge graph embedding with entity-wise top-k sparsification,”Knowledge-Based Systems, p. 114147, 2025

    2025

  26. [35]

    Fast approximate knn graph construc- tion for high dimensional data via recursive lanczos bisection

    J. Chen, H.-r. Fang, and Y . Saad, “Fast approximate knn graph construc- tion for high dimensional data via recursive lanczos bisection.”Journal of Machine Learning Research, vol. 10, no. 9, 2009

    2009

  27. [36]

    Lightgcn: Simplifying and powering graph convolution network for recommenda- tion,

    X. He, K. Deng, X. Wang, Y . Li, Y . Zhang, and M. Wang, “Lightgcn: Simplifying and powering graph convolution network for recommenda- tion,” inProceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval, 2020, pp. 639– 648

    2020

  28. [37]

    A content-driven micro-video recommendation dataset at scale,

    Y . Ni, Y . Cheng, X. Liu, J. Fu, Y . Li, X. He, Y . Zhang, and F. Yuan, “A content-driven micro-video recommendation dataset at scale,”arXiv preprint arXiv:2309.15379, 2023

    2023 arXiv

  29. [38]

    Self-supervised learning for multimedia recommendation,

    Z. Tao, X. Liu, Y . Xia, X. Wang, L. Yang, X. Huang, and T.-S. Chua, “Self-supervised learning for multimedia recommendation,”IEEE Transactions on Multimedia, vol. 25, pp. 5107–5116, 2022

    2022

  30. [39]

    Gume: Graphs and user modalities enhancement for long-tail multimodal rec- ommendation,

    G. Lin, M. Zhen, D. Wang, Q. Long, Y . Zhou, and M. Xiao, “Gume: Graphs and user modalities enhancement for long-tail multimodal rec- ommendation,” inProceedings of the 33rd ACM International Confer- ence on Information and Knowledge Management, 2024, pp. 1400–1409

    2024

  31. [40]

    Mmrec: Simplifying multimodal recommendation,

    X. Zhou, “Mmrec: Simplifying multimodal recommendation,” inPro- ceedings of the 5th ACM International Conference on Multimedia in Asia Workshops, 2023, pp. 1–2

    2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.