Pith. sign in

REVIEW 4 major objections 5 minor 52 references

Leveraging Multimodal Data and Side Users for Diffusion Cross-Domain Recommendation

T0 review · 4 major / 5 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 The paper claims that adding side users and multimodal LLM features to a diffusion-based cross-domain recommender improves cold-start accuracy, reporting the best MAE, RMSE, and NDCG among eight baselines on Amazon movie-music…

desk verdict Solid combination paper with a real but modest contribution; the abstract overclaims and the side-user distribution assumption needs testing, but it deserves serious peer review. read the letter →

arxiv 2507.04000 v1 pith:ITLKGXWF submitted 2025-07-05 cs.IR cs.AI

Fan Zhang , Jinpeng Chen , Huan Li , Senzhang Wang , Yuan Cao , Kaimin Wei , JianXiang He , Feifei Kou
show 1 more author
Jinqing Wang
This is my paper · ORCID
classification cs.IRcs.AI
keywords multimodalrecommendationcross-domaincold-startusersdiffusionmodellargelanguagemodelssidefeatureextractionAmazondataset
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

Cross-domain recommendation tries to give cold-start users useful suggestions in a target domain by transferring knowledge from an auxiliary domain where they have history. MuSiC argues that two things are being wasted: multimodal item content (images and text) that could align the two domains semantically, and 'side users' who interact only in the target domain and could teach the model that domain's embedding distribution. The proposed pipeline extracts user and item features from a multimodal large language model without fine-tuning, then trains a diffusion model in two stages: side users are reconstructed unconditionally to learn the target-domain distribution, while overlapping users are reconstructed conditioned on auxiliary features to learn the transfer pattern. Cold-start users get target-domain vectors by running the reverse diffusion process conditioned on their auxiliary features, and recommendations are dot products with target items. The paper reports that this approach outperforms the eight compared baselines on three Amazon domain pairs, with the largest gains in dual cold-start settings where both users and items are new.

What carries the argument

The load-bearing object is the cross-domain diffusion module built around a vectorized U-Net, which adapts Denoising Diffusion Probabilistic Models from image grids to 32-dimensional feature vectors. The forward process adds Gaussian noise to target-domain user feature vectors; the reverse process denoises them, optionally conditioned on auxiliary-domain feature vectors. Two-stage training is the crux: the unconditional reconstruction loss on side users (Eq. 11) teaches the model the target domain's feature distribution, while the conditional joint loss on overlapping users (Eqs. 10, 12, 13) teaches the cross-domain transformation and rating compatibility, balanced by weight $\lambda$. Feature extraction supplies the inputs: a multimodal large language model with prompt learning and a first-last-avg hidden-layer representation, passed through a tanh MLP for items, and a text-only LLM for users via their review text.

What would settle it

Construct a controlled split where side users have deliberately skewed tastes (for example, only niche genres) while cold-start users have mainstream tastes, then check whether generated cold-start vectors drift toward the side-user distribution; if MAE/RMSE degrades relative to the unskewed split, the side-user prior is doing the work. A simpler check: hold out a few cold-start users' real target-domain vectors and compare the diffusion-generated vectors' distributional distance to side users versus to true cold-start users.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that the cold-start cross-domain recommendation problem can be split into two learnable subproblems rather than one mapping: estimating the target domain's embedding distribution, and estimating the cross-domain transformation. It claims that the first subproblem is best learned from side users--target-only users who provide abundant samples of what target-domain feature vectors look like--and the second from overlapping users, whose auxiliary features pair with known target features. A single diffusion model with a vectorized U-Net carries both jobs: during training it reconstructs side users without conditioning guidance and overlapping users with auxiliary-domain conditioning; during inference it denoises Gaussian noise into a target-domain feature vector for a cold-start user, guided by that user's auxiliary features. The model also claims that multimodal features extracted from item images and text via an MLLM make the auxiliary and target representations semantically comparable before diffusion starts. The reported experiments on Amazon movie, music, and book subsets show lower MAE and RMSE and higher NDCG@20 than the eight baselines, including in the harder dual cold-start setting.

Load-bearing premise

The pivotal untested premise is that side users' target-domain feature distribution matches the distribution that cold-start users would occupy, so reconstructing side users in stage one teaches the prior needed for cold-start generation.

Editorial extensions

If this is right

  • Cold-start users with only auxiliary-domain history receive generated target-domain vectors, so recommendations in the target domain do not require any target-domain interactions.
  • Because side users supply the target-domain distribution signal, the method should remain effective when overlapping users are scarce, which the experiments test by varying the held-out proportion $\beta$ from 20% to 80%.
  • Dual cold-start scenarios--unseen users and unseen items--benefit because item vectors come from multimodal content rather than interaction histories.
  • The two-stage ordering means the model first learns what target vectors look like, then learns how auxiliary vectors change, so the auxiliary-to-target mapping is built on top of an accurate target prior.
  • Removing the MLLM feature extractor degrades performance, indicating that the content-based alignment contributes beyond the diffusion module.

Reading between the lines

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

  • If the side-user distribution assumption holds, then recruiting more target-only users should improve cold-start generation even when overlap is fixed; this is testable by adding a larger side-user pool without changing overlapping users.
  • The same two-stage recipe could be applied to more than two domains: side users from each target domain would let a shared diffusion model learn per-domain priors while overlapping users connect the domains.
  • The comparison 'w/o MLLM' still uses the same task setup, so it isolates the MLLM's contribution only partially--the pretrained extractor's general language knowledge, not multimodal content per se, may drive part of the gain.
  • The generated feature vectors are only as reliable as the rating-prediction head; the dot-product scoring could be replaced by a learned head if the 32-dimensional space is not linearly preference-compatible.
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 MuSiC, a cross-domain recommendation method for cold-start users. It uses a multimodal large language model (MiniCPM-V) and an LLM (Llama3-8B) to extract item and user features, then trains a two-stage diffusion model: first reconstructing target-domain features of side users (users active only in the target domain) without conditional guidance, and then jointly optimizing diffusion reconstruction and rating prediction for overlapping users with auxiliary-domain features as conditions. At inference, cold-start user target-domain feature vectors are generated by iterative denoising, and recommendations are made by dot product with target item features. Experiments on three Amazon domain pairs compare against eight baselines with MAE, RMSE, and NDCG@20 at three cold-start ratios, plus ablations and a dual cold-start setting with cold-start items.

Significance. If the claims were fully supported, the paper would make a useful contribution: it is, to my knowledge, the first CDR method to explicitly train a diffusion model on side users to learn the target-domain feature distribution, and the use of frozen MLLM/LLM feature extractors is a practical way to leverage multimodal content without fine-tuning. Strengths include a public code link, a standard DDPM formulation, three task setups, and an additional dual cold-start evaluation. However, the current evidence does not establish the headline claims. The abstract's statement that MuSiC 'significantly outperforms all selected baselines' is contradicted by a result in Table 2, and no statistical significance tests or variance estimates are reported anywhere. The load-bearing assumption that side-user features are distributionally matched to overlapping and cold-start users is untested. The contribution is promising but requires substantial additional verification.

major comments (4)
  1. [Abstract; Table 2] The abstract states that MuSiC 'significantly outperforms all selected baselines,' but Table 2, Task 3, β=80%, NDCG shows MuSiC at 0.8546 versus CATN at 0.8616, an improvement of -0.81%. Moreover, no significance tests, confidence intervals, or standard deviations are reported in Section 4, so the word 'significantly' is unsupported even for configurations where MuSiC is numerically best. This is a central claim of the paper and must be corrected and/or substantiated with appropriate statistical testing.
  2. [§3.5.1, Eq. (11); Table 3] The side-user diffusion loss L_dm_us in Eq. (11) is a pure reconstruction loss: side-user target features f_T_us receive no rating-prediction supervision, unlike overlapping-user features, which are also trained with L_rating in Eq. (12). Because the user-feature MLP in Section 3.2 is trainable, the model is free to map side-user reviews to a subspace that is easy to denoise but not aligned with the item-feature space used in the rating function Eq. (8). The paper provides no two-sample test or calibration check between f_T_us and f_T_uo, and the w/o side ablation in Table 3 shows small, unreplicated differences (e.g., Task 2, 80%, RMSE 1.0571 vs. 1.0466; Task 1, 80%, MAE 0.7237 vs. 0.6901). The central claim that side users improve CDR therefore needs stronger evidence, such as distributional diagnostics and significance testing.
  3. [§3.3.3] The vectorized U-Net is underspecified: the text lists initialization, down-sampling, middle block, up-sampling, and output layer, but gives no layer counts, hidden dimensions, down/up-sampling ratios, residual or attention structure, or parameter count. Since this network is the denoiser f_θ in Eqs. (10) and (11), the architecture details are essential for reproducibility and for assessing whether a U-Net is necessary. Provide a full specification and an ablation against an MLP denoiser.
  4. [§4.1.4 and §4.2] All results in Tables 2-4 appear to come from a single data split with no multiple seeds. Many margins over the best baseline are small (e.g., Task 1, 20%, NDCG 0.8650 vs. 0.8614; Task 2, 20%, RMSE 1.0136 vs. 1.0313), so it is unclear whether the reported improvements are within run-to-run variability. Report mean ± standard deviation over at least five random splits and perform paired significance tests (e.g., Wilcoxon signed-rank) for the headline comparisons.
minor comments (5)
  1. [§3.3.2, Eq. (5)] Equation (5) writes p_θ(x_{t-1}|x_t) = N(x_t; μ_θ(...), Σ_θ(...)), but the Gaussian should be over x_{t-1}, not x_t; this is presumably a typographical error.
  2. [§4.1.3] The description of the w/o MLLM ablation is confusing: 'remove the multimodal components in MuSiC and replace the user and item feature vectors with those used in the baselines' means the ablation changes both the feature extractor and the input features, so the performance drop cannot be attributed solely to removing the MLLM.
  3. [Figure 2] The labels in Figure 2 are garbled in the provided text (e.g., 'r r u u c c v v i i T T'), making the figure difficult to read.
  4. [Figures 3 and 4] The figure captions contain 'T ask' instead of 'Task' (e.g., 'T ask 1').
  5. [Section 5] The conclusion says experiments 'prove' the method's advantages; given the lack of significance testing, a more cautious wording such as 'indicate' or 'suggest' would be appropriate.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MuSiC's evaluation is a standard supervised split, and the side-user diffusion stage is a training regularizer rather than a fitted prediction re-used at test time.

full rationale

The paper's derivation chain is self-contained. User/item features are obtained by frozen LLM/MLLM encoders followed by an MLP (Eq. 1), and the diffusion module is trained with reconstruction losses on side users (Eq. 11) and overlapping users (Eqs. 10, 12-13). Cold-start test users are held-out overlapping users; their target features are generated from auxiliary-conditioned denoising and scored with a dot product (Eq. 8). No fitted constant or test-set quantity is fed back into the generator, so the reported MAE/RMSE/NDCG numbers are not predictions by construction. The main novelty claim — side users improve the learned target-domain distribution — is an empirically testable assumption, and the w/o side ablation (Table 3) is the relevant evidence; whether it is conclusive is a correctness concern, not a circularity. The only apparent self-citation, [2] EMPNet, appears as one of several generic CDR references in the introduction and is not load-bearing for any uniqueness or modeling choice. No uniqueness theorem is invoked, and no external result is renamed or re-derived.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The SOTA performance claim depends on hyperparameter choices (loss weight lambda, diffusion steps, dimension), on the unverified assumption that side-user distributions match cold-start user distributions, and on the transfer of the DDPM framework from images to low-dimensional feature vectors.

free parameters (4)
  • Diffusion reconstruction loss weight lambda = 0.2-0.6 range, chosen by grid search; per-task value not reported
    Balances diffusion loss and rating loss for overlapping users in Eq. 13 and affects the final generated vectors; the paper does not report the selected lambda per task or seed.
  • Diffusion steps T = 10
    Selected from {2,5,10,20,50} via grid search; Figure 3 shows T=10 gives best RMSE/NDCG on most settings.
  • Feature vector dimension = 32
    Fixed design choice that determines the dot-product rating scale and the capacity of all embeddings.
  • Data filtering thresholds = >=20 reviews per user, >=10 words per review
    Ad hoc preprocessing choices in Section 4.1.1 that change dataset size and may bias results toward heavy users and richer item metadata.
assumptions (4)
  • domain assumption Feature vectors extracted from auxiliary-domain reviews and target-domain item metadata lie in a space where dot products approximate user-item ratings.
    The rating function in Eq. 8 and the training objective Eq. 12 assume this linear compatibility.
  • domain assumption Side users' target-domain feature distribution matches the target-domain distribution cold-start users would have, so unsupervised reconstruction of side users improves cold-start generation.
    This is the premise of the first training stage (Eq. 11) and is not empirically verified in the paper.
  • ad hoc to paper Standard DDPM formulation (Eq. 2-6) transfers from image generation to low-dimensional user feature vectors.
    The paper adapts DDPM without derivation; the vectorized U-Net is a new architecture whose design is not justified by theory.
  • domain assumption The multimodal item features (image plus text) and user review features are informative for preference prediction.
    Section 3.2 relies on this for both feature extraction and downstream ratings.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Leveraging Multimodal Data and Side Users for Diffusion Cross-Domain Recommendation." pith.science (2026). https://pith.science/paper/ITLKGXWF

@misc{pith2026250704000,
  author       = {Pith},
  title        = {Pith review of: Leveraging Multimodal Data and Side Users for Diffusion Cross-Domain Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ITLKGXWF}},
  note         = {Machine review of arXiv:2507.04000}
}
read the original abstract

Cross-domain recommendation (CDR) aims to address the persistent cold-start problem in Recommender Systems. Current CDR research concentrates on transferring cold-start users' information from the auxiliary domain to the target domain. However, these systems face two main issues: the underutilization of multimodal data, which hinders effective cross-domain alignment, and the neglect of side users who interact solely within the target domain, leading to inadequate learning of the target domain's vector space distribution. To address these issues, we propose a model leveraging Multimodal data and Side users for diffusion Cross-domain recommendation (MuSiC). We first employ a multimodal large language model to extract item multimodal features and leverage a large language model to uncover user features using prompt learning without fine-tuning. Secondly, we propose the cross-domain diffusion module to learn the generation of feature vectors in the target domain. This approach involves learning feature distribution from side users and understanding the patterns in cross-domain transformation through overlapping users. Subsequently, the trained diffusion module is used to generate feature vectors for cold-start users in the target domain, enabling the completion of cross-domain recommendation tasks. Finally, our experimental evaluation of the Amazon dataset confirms that MuSiC achieves state-of-the-art performance, significantly outperforming all selected baselines. Our code is available: https://anonymous.4open.science/r/MuSiC-310A/.

Figures

Figures reproduced from arXiv: 2507.04000 by the authors.

Figure 1
Figure 1. (a) Existing CDR methods typically learn cross [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall architecture of MuSiC. Firstly, we utilize the user feature extraction module in the auxiliary and target [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Impact of diffusion steps 𝑇 on RMSE and NDCG for different tasks and 𝛽 values. On the left of [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Parameter sensitivity analysis on activation functions. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 33 canonical work pages

  1. [1]

    Zhicheng An, Zhexu Gu, Li Yu, Ke Tu, Zhengwei Wu, Binbin Hu, Zhiqiang Zhang, Lihong Gu, and Jinjie Gu. 2024. DDCDR: A Disentangle-based Distillation Framework for Cross-Domain Recommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 4764–4773

    work page 2024

  2. [2]

    Jinpeng Chen, Fan Zhang, Huan Li, Hua Lu, Xiongnan Jin, Kuien Liu, Hongjun Li, and Yongheng Wang. 2024. EMPNet: An extract-map-predict neural network architecture for cross-domain recommendation. World Wide Web 27, 2 (2024), 12

    work page 2024

  3. [3]

    Jingyu Chen, Lilin Zhang, and Ning Yang. 2024. Improving Adversarial Robust- ness for Recommendation Model via Cross-Domain Distributional Adversarial Training. In Proceedings of the 18th ACM Conference on Recommender Systems . 278–286

    work page 2024

  4. [4]

    Xiaolong Chen, Yifan Song, and Jing Tang. 2024. Link Recommendation to Augment Influence Diffusion with Provable Guarantees. In Proceedings of the ACM on Web Conference 2024. 2509–2518

    work page 2024

  5. [5]

    Yingjun Dai, Ahmed El-Roby, Elmira Adeeb, and Vivek Thaker. 2025. OmniMatch: Overcoming the Cold-Start Problem in Cross-Domain Recommendations using Auxiliary Reviews. In Proceedings 28th International Conference on Extending Database Technology, EDBT 2025. 80–91

    work page 2025

  6. [6]

    Liwei Deng, Yan Zhao, Yue Cui, Yuyang Xia, Jin Chen, and Kai Zheng. 2024. Task Recommendation in Spatial Crowdsourcing: A Trade-Off Between Diversity and Coverage. In 2024 IEEE 40th International Conference on Data Engineering (ICDE) . IEEE, 276–288

    work page 2024

  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. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 687–697

    work page 2024

  8. [8]

    Wenjing Fu, Zhaohui Peng, Senzhang Wang, Yang Xu, and Jin Li. 2019. Deeply fusing reviews and contents for cold start users in cross-domain recommendation systems. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 33. 94–101

    work page 2019

Show all 52 references
  1. [9]

    Yuqi Gong, Xichen Ding, Yehui Su, Kaiming Shen, Zhongyi Liu, and Guannan Zhang. 2023. An Unified Search and Recommendation Foundation Model for Cold-Start Scenario. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management . 4595–4601

    2023

  2. [10]

    Lei Guo, Ziang Lu, Junliang Yu, Quoc Viet Hung Nguyen, and Hongzhi Yin. 2024. Prompt-enhanced Federated Content Representation Learning for Cross-domain Recommendation. In Proceedings of the ACM on Web Conference 2024. 3139–3149

    2024

  3. [11]

    Zhiqiang Guo, Jianjun Li, Guohui Li, Chaoyang Wang, Si Shi, and Bin Ruan. 2024. LGMRec: Local and Global Graph Learning for Multimodal Recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 38. 8454–8462

    2024

  4. [12]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems 33 (2020), 6840–6851

    2020

  5. [13]

    Yupeng Hou, Jiacheng Li, Zhankui He, An Yan, Xiusi Chen, and Julian McAuley

  6. [14]

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

  7. [15]

    Yangqin Jiang, Yuhao Yang, Lianghao Xia, and Chao Huang. 2024. Diffkg: Knowl- edge graph diffusion model for recommendation. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining . 313–321

    2024

  8. [16]

    In Proceedings of the 32nd ACM International Conference on Multimedia

    DiffMM: Multi-Modal Diffusion Model for Recommendation. In Proceedings of the 32nd ACM International Conference on Multimedia . 7591–7599

  9. [17]

    Bohan Li, Hao Zhou, Junxian He, Mingxuan Wang, Yiming Yang, and Lei Li. 2020. On the Sentence Embeddings from Pre-trained Language Models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 9119–9130

    2020

  10. [18]

    SeongKu Kang, Junyoung Hwang, Dongha Lee, and Hwanjo Yu. 2019. Semi- supervised learning for cross-domain recommendation to cold-start users. In Proceedings of the 28th ACM international conference on information and knowledge management. 1563–1572

    2019

  11. [19]

    Jing Long, Guanhua Ye, Tong Chen, Yang Wang, Meng Wang, and Hongzhi Yin. 2024. Diffusion-Based Cloud-Edge-Device Collaborative Learning for Next POI Recommendations. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 2026–2036. Conference...

    2024

  12. [20]

    Zhenyang Li, Fan Liu, Yinwei Wei, Zhiyong Cheng, Liqiang Nie, and Mohan Kankanhalli. 2024. Attribute-driven Disentangled Representation Learning for Multimodal Recommendation. In Proceedings of the 32nd ACM International Conference on Multimedia. 9660–9669

    2024

  13. [21]

    Tong Man, Huawei Shen, Xiaolong Jin, and Xueqi Cheng. 2017. Cross-domain recommendation: An embedding and mapping approach.. In IJCAI, Vol. 17. 2464– 2470

    2017

  14. [22]

    Haokai Ma, Yimeng Yang, Lei Meng, Ruobing Xie, and Xiangxu Meng. 2024. Multimodal Conditioned Diffusion Model for Recommendation. In Companion Proceedings of the ACM on Web Conference 2024 . 1733–1740

    2024

  15. [23]

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-net: Convolu- tional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceeding...

    2015

  16. [24]

    Yifang Qin, Hongjun Wu, Wei Ju, Xiao Luo, and Ming Zhang. 2023. A diffusion model for poi recommendation. ACM Transactions on Information Systems 42, 2 (2023), 1–27

    2023

  17. [25]

    Ajit P Singh and Geoffrey J Gordon. 2008. Relational learning via collective matrix factorization. In Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining . 650–658

    2008

  18. [26]

    Yu Shang, Chen Gao, Jiansheng Chen, Depeng Jin, and Yong Li. 2024. Improving Item-side Fairness of Multimodal Recommendation via Modality Debiasing. In Proceedings of the ACM on Web Conference 2024 . 4697–4705

    2024

  19. [27]

    Caiqi Sun, Jiewei Gu, BinBin Hu, Xin Dong, Hai Li, Lei Cheng, and Linjian Mo

  20. [28]

    Zijian Song, Wenhan Zhang, Lifang Deng, Jiandong Zhang, Zhihua Wu, Kaigui Bian, and Bin Cui. 2024. Mitigating Negative Transfer in Cross-Domain Rec- ommendation via Knowledge Transferability Enhancement. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery a...

    2024

  21. [29]

    Li Wang, Lei Sang, Quangui Zhang, Qiang Wu, and Min Xu. 2024. A privacy- preserving framework with multi-modal data for cross-domain recommendation. Knowledge-Based Systems 304 (2024), 112529

    2024

  22. [30]

    Tianxin Wang, Fuzhen Zhuang, Zhiqiang Zhang, Daixin Wang, Jun Zhou, and Qing He. 2021. Low-dimensional alignment for cross-domain recommendation. In Proceedings of the 30th ACM international conference on information & knowledge management. 3508–3512

    2021

  23. [31]

    Rui Tang, Cheng Yang, and Yuxuan Wang. 2023. A Cross-Domain Multimodal Supervised Latent Topic Model for Item Tagging and Cold-Start Recommendation. IEEE MultiMedia 30, 3 (2023), 48–62

    2023

  24. [32]

    Yuhao Wang, Ziru Liu, Yichao Wang, Xiangyu Zhao, Bo Chen, Huifeng Guo, and Ruiming Tang. 2024. Diff-MSR: A Diffusion Model Enhanced Paradigm for Cold-Start Multi-Scenario Recommendation. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining . 779–787

    2024

  25. [33]

    Yuhao Wang, Yichao Wang, Zichuan Fu, Xiangyang Li, Wanyu Wang, Yuyang Ye, Xiangyu Zhao, Huifeng Guo, and Ruiming Tang. 2024. Llm4msr: An llm- enhanced paradigm for multi-scenario recommendation. In Proceedings of the 33rd ACM International Conference on Information and Knowled...

    2024

  26. [34]

    Wenjie Wang, Yiyan Xu, Fuli Feng, Xinyu Lin, Xiangnan He, and Tat-Seng Chua

  27. [35]

    In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Diffusion recommender model. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval . 832–841

  28. [36]

    Cheng Wu, Chaokun Wang, Jingcao Xu, Ziwei Fang, Tiankai Gu, Changping Wang, Yang Song, Kai Zheng, Xiaowei Wang, and Guorui Zhou. 2023. Instant Representation Learning for Recommendation over Large Dynamic Graphs. In 2023 IEEE 39th International Conference on Data Engineering (...

    2023

  29. [37]

    Zihao Wu, Xin Wang, Hong Chen, Kaidong Li, Yi Han, Lifeng Sun, and Wenwu Zhu. 2023. Diff4rec: Sequential recommendation with curriculum-scheduled diffusion augmentation. In Proceedings of the 31st ACM International Conference on Multimedia. 9329–9335

    2023

  30. [38]

    Yinwei Wei, Xiang Wang, Liqiang Nie, Xiangnan He, and Tat-Seng Chua. 2020. Graph-refined convolutional network for multimedia recommendation with implicit feedback. In Proceedings of the 28th ACM international conference on multimedia. 3541–3549

    2020

  31. [39]

    Yinwei Wei, Xiang Wang, Liqiang Nie, Xiangnan He, Richang Hong, and Tat-Seng Chua. 2019. MMGCN: Multi-modal graph convolution network for personalized recommendation of micro-video. In Proceedings of the 27th ACM international conference on multimedia. 1437–1445

    2019

  32. [40]

    Wenhao Yang, Yingchun Jian, Yibo Wang, Shiyin Lu, Lei Shen, Bing Wang, Haihong Tang, and Lijun Zhang. 2024. Not All Embeddings are Created Equal: Towards Robust Cross-domain Recommendation via Contrastive Learning. In Proceedings of the ACM on Web Conference 2024 . 3195–3206

    2024

  33. [41]

    Wei Yang, Jie Yang, and Yuan Liu. 2023. Multimodal Optimal Transport Knowl- edge Distillation for Cross-domain Recommendation. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management . 2959– 2968

    2023

  34. [42]

    Wujiang Xu, Shaoshuai Li, Mingming Ha, Xiaobo Guo, Qiongxu Ma, Xiaolei Liu, Linxun Chen, and Zhenfeng Zhu. 2023. Neural node matching for multi-target cross domain recommendation. In 2023 IEEE 39th International Conference on Data Engineering (ICDE). IEEE, 2154–2166

    2023

  35. [43]

    Yuner Xuan. 2024. Diffusion Cross-domain Recommendation. arXiv:2402.02182

    2024 arXiv

  36. [44]

    Cheng Zhao, Chenliang Li, Rong Xiao, Hongbo Deng, and Aixin Sun. 2020. CATN: Cross-domain recommendation for cold-start users via aspect transfer network. In Proceedings of the 43rd international ACM SIGIR conference on research and development in information retrieval. 229–238

    2020

  37. [45]

    Chuang Zhao, Hongke Zhao, Ming He, Jian Zhang, and Jianping Fan. 2023. Cross- domain recommendation via user interest alignment. In Proceedings of the ACM Web Conference 2023. 887–896

    2023

  38. [46]

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

    2024

  39. [47]

    Yixin Zhang, Yong Liu, Hao Xiong, Yi Liu, Fuqiang Yu, Wei He, Yonghui Xu, Lizhen Cui, and Chunyan Miao. 2023. Cross-domain disentangled learning for e-commerce live streaming recommendation. In 2023 IEEE 39th International Conference on Data Engineering (ICDE) . IEEE, 2955–2968

    2023

  40. [48]

    Yongchun Zhu, Zhenwei Tang, Yudan Liu, Fuzhen Zhuang, Ruobing Xie, Xu Zhang, Leyu Lin, and Qing He. 2022. Personalized transfer of user preferences for cross-domain recommendation. In Proceedings of the fifteenth ACM international conference on web search and data mining . 150...

    2022

  41. [50]

    Chuang Zhao, Hongke Zhao, Xiaomeng Li, Ming He, Jiahui Wang, and Jianping Fan. 2023. Cross-domain recommendation via progressive structural alignment. IEEE Transactions on Knowledge and Data Engineering (2023)

    2023

  42. [51]

    Jujia Zhao, Wang Wenjie, Yiyan Xu, Teng Sun, Fuli Feng, and Tat-Seng Chua. 2024. Denoising diffusion recommender model. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval . 1370–1379

    2024

  43. [2023]

    In Proceedings of the AAAI Conference on Artificial Intelligence , Vol

    REMIT: reinforced multi-interest transfer for cross-domain recommen- dation. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 37. 9900–9908

  44. [2024]

    arXiv:2403.03952

    Bridging Language and Items for Retrieval and Recommendation. arXiv:2403.03952

    arXiv

Pith tools

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