Pith. sign in

REVIEW 4 major objections 4 minor 43 references

RaMen: Multi-Strategy Multi-Modal Learning for Bundle Construction

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A multimodal architecture that fuses explicit item-view attention with hypergraph intent discovery lifts bundle construction recall by up to 77 percent over the previous best model.

desk verdict RaMen is a genuinely new multi-strategy architecture for bundle construction, but its headline gains rest on inherited baselines and unreported variance, so the numbers are not yet credible as reported. read the letter →

arxiv 2507.14361 v1 pith:XIUO6YGW submitted 2025-07-18 cs.IR

Huy-Son Nguyen , Quang-Huy Nguyen , Duc-Hoang Pham , Duc-Trong Le , Hoang-Quynh Le , Padipat Sitkrongwong , Atsuhiro Takasu , Masoud Mansoury This is my paper · ORCID
classification cs.IR
keywords bundleconstructionmulti-modallearninghypergraphcontrastivegraphattentionitemco-purchasemulti-strategyrecommendersystems
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

This paper claims that automatic bundle construction improves when a model learns two complementary strategies at once: an explicit strategy that reads an item's own characteristics (text, image, and direct co-purchase relations) and an implicit strategy that discovers shared latent intents among groups of items through hyperedges. The proposed architecture, RaMen, combines these two views, reconciles them with a contrastive alignment module, and scores candidate items by inner product under a negative log-likelihood objective. On four datasets spanning fashion, music, electronics, and food, RaMen is reported to beat the previous best method CLHE by 0.83% to 77.31% in Recall@20, with the largest gains on small bundles built for specific intents. If these results hold, multi-strategy learning of this kind would give practitioners a way to construct bundles that single-view, feedback-only models miss.

What carries the argument

The load-bearing mechanism is the pairing of two representation strategies with an alignment loss. ESL is task-specific attention: a characteristic encoder runs self-attention over concatenated visual, textual, and ID embeddings of items and of bundles, while a collaborative encoder propagates attention-weighted signals over a homogeneous item graph whose edges come from thresholding the co-purchase matrix $E = X^{\top} X$. ISL introduces learnable hyperedge embeddings $W^{m}$ for each modality $m \in \{t, v\}$, forms item-hyperedge dependency matrices $F_I^{m} = M_I^{m} (W^{m})^{\top}$ and bundle-hyperedge matrices $F_B^{m} = Y (F_I^{m})^{\top}$, denoises them with Gumbel-Softmax reparameterization, and propagates embeddings over the hypergraph for $Z$ layers. MAD then pulls the ESL and ISL representations of the same item or bundle together with InfoNCE while pushing different objects apart, and retrieval scores candidate items by $\sigma_{b,i} = g_b \cdot g_i^{\top} + \phi_b \cdot \phi_i^{\top}$, trained with a negative log-likelihood objective plus the weighted contrastive and regularization terms.

What would settle it

Re-run every baseline with the same code, splits, feature files, and hyperparameter grid used for RaMen, then compare Recall@20 on Electronic and Food: if the re-run CLHE numbers land near RaMen's 0.8371 and 0.8459 rather than the inherited 0.4721 and 0.5077, the headline 77.31% and 66.61% margins are artifacts of the comparison setup rather than of the method.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that bundle representations are best learned by modeling two distinct decision-making strategies and then reconciling them. Explicit Strategy-aware Learning (ESL) applies task-specific attention over multi-modal item features and over a homogeneous item co-purchase graph, capturing what each bundle is explicitly about; Implicit Strategy-aware Learning (ISL) learns hyperedge embeddings per modality and propagates messages over the resulting hypergraph, capturing shared latent intents that no pairwise relation exposes. The Multi-strategy Alignment & Discrimination (MAD) module transfers knowledge between the two strategies with an InfoNCE contrastive loss while keeping distinct items and bundles separated in the embedding space. The reported outcome is that the combination outperforms all six considered baselines on all four datasets, with Recall@20 gains over the strongest baseline CLHE of 32.04% on POG, 0.83% on Spotify, 77.31% on Electronic, and 66.61% on Food, and the ablations attribute the largest drops to removing the collaborative encoder on dense data and to removing the characteristic encoder on sparse data.

Load-bearing premise

The reported margins over prior methods assume that the feature files for Spotify and POG and the baseline scores for Bi-LSTM, HyperGraph, Trans, TransCL, GAT, and CLHE, all inherited from the earlier CLHE study, are correct and directly comparable to RaMen's own runs, and the paper itself notes it could not reproduce the inherited Bi-LSTM result on Spotify.

Editorial extensions

If this is right

  • If the reported margins hold, feedback-only bundle construction built on user-item bipartite graphs leaves substantial performance unused; explicit item characteristics and implicit group intents are complementary signals, not optional extras.
  • Small-bundle, intent-specific domains (electronics, food, fashion outfits) stand to gain the most, since these show the largest relative jumps in Recall@20, while dense large-bundle domains such as music playlists gain less but still improve in ranking metrics.
  • The ablations give a tuning rule of thumb: weight item characteristics and textual features more heavily on sparse data, and weight collaborative signals more heavily on dense data.
  • A small number of hyperedges (4 to 8) suffices for the implicit strategy, so the hypergraph component adds latent-intent modeling at modest computational cost.

Reading between the lines

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

  • An extension the paper does not test: the same explicit/implicit split could be applied to recommending pre-existing bundles, where the explicit strategy would read bundle descriptions rather than construct them, since nothing in the mechanism requires item-set prediction as the only task.
  • The per-dataset hand-tuned co-purchase threshold that the paper flags as a limit means portability costs calibration; an adaptive edge-filtering rule would be a direct stress test of whether the homogeneous item graph or the threshold choice carries the collaborative gains.
  • The hyperedge representation should give the model a cold-start advantage for items with no co-purchase history, because textual and visual affinity to learned intents could place such items in bundles before any collaborative evidence exists; the paper does not evaluate this regime.
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 / 4 minor

Summary. The paper proposes RaMen, a bundle construction framework that combines Explicit Strategy-aware Learning (ESL), which encodes item characteristics and collaborative signals via attention and graph propagation, with Implicit Strategy-aware Learning (ISL), which uses learnable hyperedge embeddings and hypergraph message passing to capture latent intents. A Multi-strategy Alignment & Discrimination (MAD) module applies contrastive loss to align strategy-specific representations. Experiments on POG, Spotify, Electronic, and Food report substantial gains over CLHE and other baselines, with up to 77.31% relative Recall@20 improvement on Electronic. The paper also provides ablation studies and qualitative examples.

Significance. If the reported results hold, RaMen represents a meaningful architectural advance for bundle construction, particularly in combining multi-modal features with explicit and implicit relational modeling. The authors provide code and a thorough ablation study that isolates the contribution of each module, which strengthens internal validity. However, the central claim of state-of-the-art performance rests on an evaluation protocol that inherits baseline numbers and features from prior work without independent verification, and the statistical significance claims are not backed by reported variance. This limits the external validity of the superiority claim until the evaluation is made fully reproducible and comparable.

major comments (4)
  1. [Section 4.1.3, Table 2] The claim of statistical significance (‡, p < 0.05) is not supported by the reported evidence. The table notes that significance was 'obtained through the average performance of five runs of each model,' but no standard deviations, confidence intervals, or per-run values are provided for RaMen or for any baseline. Furthermore, the baseline numbers are taken directly from prior work [28] (and Bi-LSTM results are explicitly acknowledged as not reproduced on Spotify), so the significance test compares a five-run average against a single inherited number from a different codebase. This undermines the central superiority claim and requires either rerunning all baselines under the identical protocol or reporting full variance information.
  2. [Section 4.1.1, Section 4.1.3] The empirical comparison inherits multimodal features for Spotify and POG from [28] and uses baseline results from the same work without independent reruns. The paper does not demonstrate that the dataset splits, random masking protocol, or evaluation code are identical between the two pipelines. Given that the headline margins (e.g., 77.31% R@20 on Electronic) are computed relative to CLHE numbers from [28], any subtle difference in preprocessing or evaluation could change the comparison. The authors should either rerun all baselines under the exact same conditions or provide the exact evaluation code and data splits to make the comparison verifiable.
  3. [Section 4.1.3] The hyperparameters (ϵ, γ, β, λ1, H, L1, L2, N, Z) are described as 'empirically tuned' and 'tuned ... using grid search' per dataset, but the paper does not explicitly state that they were selected using validation data only. If test-set feedback influenced hyperparameter selection, the reported performance would be optimistically biased. A clear statement of the validation-only tuning protocol is necessary to support the claim of fair comparison.
  4. [Eq. (10), Section 3.3] The 'Gumbel-Softmax' reparameterization in Eq. (10) uses the term log θ − log(1 − θ), which is the logit of a uniform random variable, not a Gumbel variate (standard Gumbel noise is −log(−log U)). This means the dependency matrices are not actually Gumbel-Softmax distributed as claimed. While this may not change the empirical ranking, it is a technical inaccuracy in a central module and should be corrected or clarified, possibly by reworking the sampling scheme.
minor comments (4)
  1. [Eq. (4)] In the bundle-level attention equation, the rightmost term in the softmax appears to use ilde{\rho}^{(l-1)}_i rather than ilde{\rho}^{(l-1)}_b, which is inconsistent with self-attention over bundle features; this is likely a typographical error.
  2. [Introduction and Section 3.4] The module is called 'Multi-strategy Alignment & Dispersion' in the introduction and 'Multi-strategy Alignment & Discrimination' in Section 3.4; the terminology should be made consistent.
  3. [Eq. (11)] The bundle propagation formula ϕ^{m,(z+1)}_b = That{F}^m_B (\hat{F}^m_I)^⊤ ϕ^{m,(z)}_i is unclear because the subscript i on ϕ^{m,(z)}_i is not defined; it should be clarified that this aggregates over items in the bundle, or the equation should be written in matrix form.
  4. [Section 4.2] The text states that RaMen achieves 'modest yet steady gains on Spotify,' but Table 2 shows R@10 and R@20 improvements of only 0.87% and 0.83%, which are not marked as significant; this should be reflected in the wording to avoid overstating the result.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: RaMen's reported gains are empirical benchmark results from a learned model; inherited CLHE baselines and unreproduced Bi-LSTM are reproducibility caveats, not circularity.

full rationale

The paper's central claim is empirical: RaMen achieves better Recall@K and NDCG@K than CLHE and other baselines (Table 2). The model is a learned encoder over item features, a user-item co-purchase graph, and bundle-item incidence; the training objective (NLL, Eq. 16) and retrieval score (Eq. 15) are standard set-completion losses, not defined in terms of the reported metrics. No equation in Section 3 makes a predicted bundle equivalent, by construction, to a fitted input: hyperedge dependency matrices F_I and F_B are learned low-rank projections of modality features and bundle incidence (Eq. 9), and the item graph E = X^T X is an input representation, not a renamed output. Per-dataset hyperparameters (epsilon, H, gamma, N) are tuned by grid search (Section 4.1.3); this is standard model selection, and the paper does not claim a parameter-free first-principles prediction. Self-citations [29,30] appear as method inspiration and hyperparameter ranges, but the superiority claim rests on the experiments, not on those citations. The main caveats are reproducibility-related, not circular: Section 4.1.3 explicitly admits Bi-LSTM results were not reproduced, baseline numbers are inherited from [28], and the significance markers are computed from five runs of RaMen against inherited baseline values. These are legitimate external validation risks, but they do not make the derivation circular.

Assumptions & free parameters 11 free parameters · 5 assumptions · 1 invented entities

The model's performance depends on a large set of hand-tuned hyperparameters (epsilon, gamma, beta, lambda_1, H, layer counts, tau) and on domain assumptions about the validity of the autoencoder formulation, the co-purchase graph, and the inherited evaluation setup. No invented entity with independent empirical support is introduced beyond the interpretive construct of latent intents.

free parameters (11)
  • epsilon (co-purchase graph threshold) = 5 (POG), 450 (Spotify), 1 (Food/Electronic)
    Discretizes the item co-purchase matrix into the item homogeneous graph; tuned per dataset based on interaction distribution (Sections 3.1.3, 4.1.3).
  • gamma (strategy balance) = tuned in {0.1..0.9}
    Balances characteristic and collaborative encoders in Eqs. (7)-(8); sensitive hyperparameter per Figure 2.
  • beta (residual weight) = tuned in {0.1..0.9}
    Controls residual connection strength in Collaborative Strategy Encoder, Eq. (6).
  • lambda_1 (contrastive weight) = tuned in {0.1..0.9}
    Weights the MAD contrastive loss in total objective Eq. (17).
  • H (number of hyperedges) = chosen from {4,8,16,32,64}
    Latent intent granularity in ISL; low values (4 or 8) preferred per Figure 2.
  • L1, L2 (attention layers in ESL) = each in {1..5}
    Depth of characteristic and bundle attention encoders, values empirically explored.
  • N (collaborative attention contexts) = in {1..5}; optimal 2 (Electronic), 4 (POG)
    Number of propagation steps in item-item graph attention; deeper stacks oversmooth (Figure 3).
  • Z (hypergraph layers) = in {1..5}
    Number of hypergraph propagation layers in ISL, Eq. (11).
  • tau (Gumbel-Softmax temperature) = 0.2
    Temperature for dependency matrix reparameterization, Eq. (10), 'empirically selected'.
  • lambda_2 (regularization) = 1e-5
    Weight decay on trainable parameters, Eq. (17).
  • embedding size, batch size, learning rate = 64, 1024, 1e-3
    Standard configuration inherited from related works, Section 4.1.3.
assumptions (5)
  • domain assumption Bundle construction can be formulated as an autoencoder task where a full bundle is encoded and then reconstructed from its items.
    Section 3.1.1 states training complies with an auto-encoder approach inherited from CLHE; the central evaluation depends on this formulation.
  • domain assumption Co-purchase counts X^T X, discretized with threshold epsilon, capture meaningful direct item-item collaborative relations.
    Section 3.1.3 constructs the item homogeneous graph by thresholding the co-purchase matrix; this graph is the basis of the Collaborative Strategy Encoder.
  • ad hoc to paper Learnable hyperedge embeddings with Gumbel-Softmax dependency matrices uncover shared latent intents among item groups.
    Section 3.3 introduces hypergraph message passing as the Implicit Strategy-aware Learning; the interpretive claim about 'hidden intents' is not independently validated.
  • domain assumption InfoNCE contrastive loss between strategy-specific representations improves cross-strategy alignment and discrimination.
    Section 3.4 uses MAD loss; this is a standard assumption in contrastive learning but not established for this specific task.
  • domain assumption Evaluation protocol and pre-extracted multimodal features inherited from CLHE are valid for fair comparison.
    Section 4.1.1 states features for Spotify and POG are inherited from baseline work [28]; baselines are 'acknowledged available results' from the same work, so the comparison depends on their correctness.
invented entities (1)
  • Latent bundle intents encoded as learnable hyperedge embeddings W_m
    purpose: To model implicit shared attributes among groups of items in ISL.
    The hyperedges are learned from the same training data and their 'intent' interpretation is post hoc; no external validation (e.g., human labels or behavioral prediction) is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RaMen: Multi-Strategy Multi-Modal Learning for Bundle Construction." pith.science (2026). https://pith.science/paper/XIUO6YGW

@misc{pith2026250714361,
  author       = {Pith},
  title        = {Pith review of: RaMen: Multi-Strategy Multi-Modal Learning for Bundle Construction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XIUO6YGW}},
  note         = {Machine review of arXiv:2507.14361}
}
read the original abstract

Existing studies on bundle construction have relied merely on user feedback via bipartite graphs or enhanced item representations using semantic information. These approaches fail to capture elaborate relations hidden in real-world bundle structures, resulting in suboptimal bundle representations. To overcome this limitation, we propose RaMen, a novel method that provides a holistic multi-strategy approach for bundle construction. RaMen utilizes both intrinsic (characteristics) and extrinsic (collaborative signals) information to model bundle structures through Explicit Strategy-aware Learning (ESL) and Implicit Strategy-aware Learning (ISL). ESL employs task-specific attention mechanisms to encode multi-modal data and direct collaborative relations between items, thereby explicitly capturing essential bundle features. Moreover, ISL computes hyperedge dependencies and hypergraph message passing to uncover shared latent intents among groups of items. Integrating diverse strategies enables RaMen to learn more comprehensive and robust bundle representations. Meanwhile, Multi-strategy Alignment & Discrimination module is employed to facilitate knowledge transfer between learning strategies and ensure discrimination between items/bundles. Extensive experiments demonstrate the effectiveness of RaMen over state-of-the-art models on various domains, justifying valuable insights into complex item set problems.

Figures

Figures reproduced from arXiv: 2507.14361 by the authors.

Figure 1
Figure 1. Overall architecture of our proposed model RaMen. with an auto-encoder approach [28], where the entire items within the bundle are considered as the input, and the same set should be predicted as the output. 3.1.2 Semantic Information Extraction. Inherited from [24, 28], the textual and visual features of each item are derived from large-scale multi-modal feature extractors, repre￾sented as {mt i ∈ R dmt , mv i ∈ R … view at source ↗
Figure 2
Figure 2. Impact of γ, H on RaMen’s performance. where we modeled each component, and tuned it using grid search. These observations are clarified with the aim of providing the essen￾tial value range and insights for extending further work [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Impact of hyperparameter N on RaMen’s performance. In particular, the fluctuation of γ further reinforces the critical role of both CbSE and CrSE on different data domains, as discussed. As γ increases, the performance of RaMen initially improves, reaches an optimal value, and then declines across both datasets. This behavior can be attributed to the role of γ as a balancing parameter, which reg￾ulates the relative … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Practical studies of top-K item candidates of RaMen compared to CLHE across the POG and Electronic dataset. The green box surrounding the item’s image denotes the item that the model has correctly predicted. ture similarity. This approach often results in the predictio…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 40 canonical work pages

  1. [28]

    Y . Ma, X. Liu, Y . Wei, Z. Tao, X. Wang, and T.-S. Chua. Leveraging multimodal features and item-level user feedback for bundle construc- tion. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining, pages 510–519, 2024

    work page 2024

  2. [1]

    Brody, U

    S. Brody, U. Alon, and E. Yahav. How attentive are graph attention networks? In International Conference on Learning Representations , 2022

    work page 2022

  3. [2]

    Bui, H.-S

    T.-N. Bui, H.-S. Nguyen, C.-V . N. Thi, H.-Q. Le, and D.-T. Le. Bridge: Bundle recommendation via instruction-driven generation. arXiv preprint arXiv:2412.18092, 2024

    arXiv 2024

  4. [3]

    Bui, H.-S

    T.-N. Bui, H.-S. Nguyen, C.-V . T. Nguyen, H.-Q. Le, and D.-T. Le. Per- sonalized diffusion model reshapes cold-start bundle recommendation. In Companion Proceedings of the ACM on Web Conference 2025, pages 3088–3091, 2025

    work page 2025

  5. [4]

    Chang, C

    J. Chang, C. Gao, X. He, D. Jin, and Y . Li. Bundle recommendation and generation with graph neural networks. IEEE Transactions on Knowl- edge and Data Engineering, 35(3):2326–2340, 2021

    work page 2021

  6. [5]

    C.-W. Chen, P. Lamere, M. Schedl, and H. Zamani. Recsys challenge 2018: Automatic music playlist continuation. InProceedings of the 12th ACM Conference on Recommender Systems, pages 527–528, 2018

    work page 2018

  7. [6]

    L. Chen, Y . Liu, X. He, L. Gao, and Z. Zheng. Matching user with item set: Collaborative bundle recommendation with deep attention network. In IJCAI, pages 2095–2101, 2019

    work page 2019

  8. [7]

    W. Chen, P. Huang, J. Xu, X. Guo, C. Guo, F. Sun, C. Li, A. Pfadler, H. Zhao, and B. Zhao. Pog: personalized outfit generation for fashion recommendation at alibaba ifashion. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data min- ing, pages 2662–2670, 2019

    work page 2019

Show all 43 references
  1. [8]

    Q. Deng, K. Wang, M. Zhao, R. Wu, Y . Ding, Z. Zou, Y . Shang, J. Tao, and C. Fan. Build your own bundle-a neural combinatorial optimization method. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2625–2633, 2021

    2021

  2. [9]

    Z. Deng, J. Li, Z. Guo, W. Liu, L. Zou, and G. Li. Multi-view multi- aspect neural networks for next-basket recommendation. In Proceed- ings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1283–1292, 2023

    2023

  3. [10]

    Y . Ding, P. Mok, Y . Ma, and Y . Bin. Personalized fashion outfit genera- tion with user coordination preference learning.Information Processing & Management, 60(5):103434, 2023

    2023

  4. [11]

    X. Du, K. Qian, Y . Ma, and X. Xiang. Enhancing item-level bundle representation for bundle recommendation. ACM Transactions on Rec- ommender Systems, 2023

    2023

  5. [12]

    Y . Fang, X. Xiao, X. Wang, and H. Lan. Customized bundle recom- mendation by association rules of product categories for online super- markets. In 2018 IEEE Third International Conference on Data Science in Cyberspace (DSC), pages 472–475. IEEE, 2018

    2018

  6. [13]

    C. Gao, Y . Zheng, N. Li, Y . Li, Y . Qin, J. Piao, Y . Quan, J. Chang, D. Jin, X. He, et al. A survey of graph neural networks for recommender systems: Challenges, methods, and directions. ACM Transactions on Recommender Systems, 1(1):1–51, 2023

    2023

  7. [14]

    Z. Guo, J. Li, G. Li, C. Wang, S. Shi, and B. Ruan. Lgmrec: Local and global graph learning for multimodal recommendation. In Pro- ceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 8454–8462, 2024

    2024

  8. [15]

    X. Han, Z. Wu, Y .-G. Jiang, and L. S. Davis. Learning fashion com- patibility with bidirectional lstms. In Proceedings of the 25th ACM international conference on Multimedia, pages 1078–1086, 2017

    2017

  9. [16]

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

    2020

  10. [17]

    R. T. Irene, C. Borrelli, M. Zanoni, M. Buccoli, and A. Sarti. Automatic playlist generation using convolutional neural networks and recurrent neural networks. In 2019 27th European signal processing conference (EUSIPCO), pages 1–5. IEEE, 2019

    2019

  11. [18]

    E. Jang, S. Gu, and B. Poole. Categorical reparametrization with gumbel-softmax. In International Conference on Learning Represen- tations (ICLR 2017). OpenReview. net, 2017

    2017

  12. [19]

    D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In Y . Bengio and Y . LeCun, editors, 3rd International Conference on Learning Representations, 2015

    2015

  13. [20]

    J. Li, D. Li, C. Xiong, and S. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888–12900. PMLR, 2022

    2022

  14. [21]

    M. Li, S. Jullien, M. Ariannezhad, and M. de Rijke. A next basket rec- ommendation reality check.ACM Transactions on Information Systems, 41(4):1–29, 2023

    2023

  15. [22]

    F. Liu, H. Chen, Z. Cheng, A. Liu, L. Nie, and M. Kankanhalli. Disen- tangled multimodal representation learning for recommendation. IEEE Transactions on Multimedia, 25:7149–7159, 2022

    2022

  16. [23]

    G. Liu, Y . Fu, G. Chen, H. Xiong, and C. Chen. Modeling buying mo- tives for personalized product bundle recommendation. ACM Transac- tions on Knowledge Discovery from Data (TKDD), 11(3):1–26, 2017

    2017

  17. [24]

    Q. Liu, J. Hu, Y . Xiao, X. Zhao, J. Gao, W. Wang, Q. Li, and J. Tang. Multimodal recommender systems: A survey.ACM Computing Surveys, 2023

    2023

  18. [25]

    X. Liu, J. Wu, Z. Tao, Y . Ma, Y . Wei, and T.-s. Chua. Fine-tuning multimodal large language models for product bundling. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V . 1, pages 848–858, 2025

    2025

  19. [26]

    Y . Ma, Y . He, A. Zhang, X. Wang, and T.-S. Chua. Crosscbr: Cross- view contrastive learning for bundle recommendation. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 1233–1241, 2022

    2022

  20. [27]

    Y . Ma, Y . He, W. Zhong, X. Wang, R. Zimmermann, and T.-S. Chua. Cirp: Cross-item relational pre-training for multimodal product bundling. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 9641–9649, 2024

    2024

  21. [29]

    Nguyen, T.-N

    H.-S. Nguyen, T.-N. Bui, L.-H. Nguyen, D.-C. Can, C.-V . T. Nguyen, D.-T. Le, and H.-Q. Le. Hhmc: a heterogeneous x homogeneous graph- based network for multimodal cross-selling recommendation. In the 15th International Conference on Knowledge and Systems Engineering, pages 1–...

    2023

  22. [30]

    Nguyen, T.-N

    H.-S. Nguyen, T.-N. Bui, L.-H. Nguyen, H. Hoang, C.-V . Thi Nguyen, H.-Q. Le, and D.-T. Le. Bundle recommendation with item-level causation-enhanced multi-view learning. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases , pages 324–341. Spri...

    2024

  23. [31]

    A. v. d. Oord, Y . Li, and O. Vinyals. Representation learning with con- trastive predictive coding. arXiv preprint arXiv:1807.03748, 2018

    2018 arXiv

  24. [32]

    Pathak, K

    A. Pathak, K. Gupta, and J. McAuley. Generating and personalizing bundle recommendations on steam. In Proceedings of the 40th Interna- tional ACM SIGIR Conference on Research and Development in Infor- mation Retrieval, pages 1073–1076, 2017

    2017

  25. [33]

    M. Sun, L. Li, M. Li, X. Tao, D. Zhang, P. Wang, and J. X. Huang. A survey on bundle recommendation: Methods, applications, and chal- lenges. arXiv preprint arXiv:2411.00341, 2024

    2024

  26. [34]

    Z. Sun, K. Feng, J. Yang, H. Fang, X. Qu, Y .-S. Ong, and W. Liu. Revis- iting bundle recommendation for intent-aware product bundling. ACM Transactions on Recommender Systems, 2(3):1–34, 2024

    2024

  27. [35]

    Z. Sun, K. Feng, J. Yang, X. Qu, H. Fang, Y .-S. Ong, and W. Liu. Adap- tive in-context learning with large language models for bundle genera- tion. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 966–976, 2024

    2024

  28. [36]

    A. Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017

    2017

  29. [37]

    Y . Wei, X. Liu, Y . Ma, X. Wang, L. Nie, and T.-S. Chua. Strategy-aware bundle recommender system. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1198–1207, 2023

    2023

  30. [38]

    L. Xia, C. Huang, Y . Xu, J. Zhao, D. Yin, and J. Huang. Hypergraph contrastive collaborative filtering. In Proceedings of the 45th Interna- tional ACM SIGIR conference on research and development in informa- tion retrieval, pages 70–79, 2022

    2022

  31. [39]

    Z. Yu, J. Li, L. Chen, and Z. Zheng. Unifying multi-associations through hypergraph for bundle recommendation. Knowledge-Based Systems, 255:109755, 2022

    2022

  32. [40]

    S. Zhao, W. Wei, D. Zou, and X. Mao. Multi-view intent disentan- gle graph networks for bundle recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 36, pages 4379– 4387, 2022

    2022

  33. [41]

    H. Zhou, X. Zhou, L. Zhang, and Z. Shen. Enhancing dyadic relations with homogeneous graphs for multimodal recommendation. In ECAI 2023, pages 3123–3130. IOS Press, 2023

    2023

  34. [42]

    Zhou and Z

    X. Zhou and Z. Shen. A tale of two graphs: Freezing and denoising graph structures for multimodal recommendation. In Proceedings of ACM International Conference on Multimedia, pages 935–943, 2023

    2023

  35. [43]

    T. Zhu, P. Harrington, J. Li, and L. Tang. Bundle recommendation in e-commerce. In Proceedings of the 37th international ACM SIGIR conference on Research & development in information retrieval, pages 657–666, 2014

    2014

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.