Exploring Preference-Guided Diffusion Model for Cross-Domain Recommendation

Dawei Yin; Hengzhu Tang; Jiawei Sheng; Li Gao; Suqi Cheng; Tingwen Liu; Xiaodong Li; Xinghua Zhang

REVIEW 3 major objections 6 minor 1 cited by

Exploring Preference-Guided Diffusion Model for Cross-Domain Recommendation

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Preference-guided diffusion, applied step by step in the reverse process, generates better target-domain user representations for cold-start users than embedding-mapping or meta-learning baselines.

desk verdict Preference-guided diffusion for cross-domain recommendation with large reported gains, but the missing code, error bars, and a weak non-diffusion control mean the central claim needs more evidence before acceptance. read the letter →

arxiv 2501.11671 v1 pith:FVKUQ2EE submitted 2025-01-20 cs.IR

Xiaodong Li , Hengzhu Tang , Jiawei Sheng , Xinghua Zhang , Li Gao , Suqi Cheng , Dawei Yin , Tingwen Liu This is my paper · ORCID

classification cs.IR

keywords cross-domainrecommendationcold-startdiffusionmodelspreferenceguidanceratingpredictionAmazonreviewdatasetclassifier-free

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

The paper proposes DMCDR, a diffusion-model approach to cross-domain recommendation for cold-start users, who have interaction history in one domain but none in another. Its central claim is that explicitly injecting a source-domain preference signal into every step of the diffusion reverse process yields a more personalized target-domain user representation than embedding-mapping or meta-learning baselines. On three Amazon cross-domain scenarios (Movie→Music, Book→Movie, Book→Music), the authors report consistent gains over the best baselines, up to 32.60% lower MAE and 13.90% lower RMSE. The paper also compares six diffusion-based variants to isolate where preference guidance should enter the diffusion pipeline, concluding that step-by-step guidance during reverse denoising is the key design choice.

What carries the argument

The preference-guided reverse process: a preference encoder (Transformer layer plus average pooling) produces the guidance signal $\boldsymbol{h}^s_i$ from source-domain interaction history; the reverse transition $p_\theta(\hat{\boldsymbol{u}}_{i,t-1} \mid \hat{\boldsymbol{u}}_{i,t}, \boldsymbol{h}^s_i)$ is an MLP that takes the corrupted representation, the guidance signal, and a diffusion-step embedding as inputs and directly predicts the clean representation $\hat{\boldsymbol{u}}_{i,0}$; classifier-free guidance blends conditional and unconditional predictions with strength $\omega$; and an exponential noise schedule controls the forward noise so that personalized information is not destroyed. The guidance is injected at every denoising step rather than only at the start or end, which the variant study suggests is the critical design decision.

What would settle it

Compare DMCDR's current inference initialization (the cold-start user's untrained target embedding) against starting the same reverse process from freshly sampled standard Gaussian noise, and also against a fixed shared random vector, across the three Amazon scenarios. If the untrained-embedding start does not yield lower MAE and RMSE than the Gaussian start, the claimed benefit of that initialization is refuted.

Watch

Extended reading notes

Core claim

DMCDR treats the target-domain user representation as a noisy signal to be denoised under the guidance of a preference encoder that summarizes the user's source-domain interaction history using a Transformer layer and average pooling. The forward process adds controlled Gaussian noise with an exponential schedule; the reverse process, parameterized by an MLP, reconstructs the user representation while receiving the preference guidance signal at every step, with classifier-free guidance controlling the strength of that signal. At inference, cold-start users start the reverse process from their untrained target-domain embedding rather than pure Gaussian noise, and the resulting representation is used for rating prediction. The authors claim that this explicit, step-by-step preference injection is what allows the model to outperform both mapping-based and meta-learning state-of-the-art methods, as well as the six diffusion variants they explore.

Load-bearing premise

Inference starts the reverse process from the cold-start user's target-domain embedding, which has never been trained on target-domain signals; the method assumes that denoising this untrained random vector under source-preference guidance beats starting from standard Gaussian noise, and that assumption is not tested by the ablations.

Editorial extensions

If this is right

If DMCDR's central claim holds, conditional diffusion becomes a viable alternative paradigm to embedding-and-mapping for cross-domain cold-start recommendation.
The step-wise injection design indicates that preference transfer benefits from being distributed across the denoising trajectory rather than applied once at the input or output.
The method retains its advantage even when only 20% of overlapping users are available for training, suggesting it is suitable for target domains with sparse interaction data.
Predicting the user representation directly instead of predicting the added noise aligns the diffusion objective with the recommendation task and appears to be an important ingredient.
The six-variant comparison suggests that noising the preference signal itself (forward-process injection) damages personalized information, so guidance should condition the reverse process only.

Reading between the lines

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

One testable extension the paper does not run: replace the cold-start user's untrained target embedding at inference with standard Gaussian noise; if performance does not degrade, the initialization is not the source of the gain.
Because the inference start is a user-specific random vector, the result may depend on the particular random draw; averaging over several random seeds for that embedding would quantify how sensitive DMCDR is to this initialization.
The exponential noise schedule with small $\eta$ keeps the forward process close to the original representation, hinting that the diffusion steps may function more like learned iterative refinement than full generative denoising; varying $\eta$ over a wider range could test this distinction.
The preference encoder uses pooled Transformer outputs, so the method treats source history as a bag of items; extending it to sequential or multi-domain preference signals could reveal whether order information further improves cold-start transfer.

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.

Desk Editor's Note

read the letter

Here’s what you should know about this one: DMCDR is a diffusion-based cross-domain recommender that injects a source-domain preference summary into the reverse process at every step, then denoises a target-domain user representation for cold-start users. The reported gains are large—22–33% MAE improvement over CDRNP and DiffCDR on three Amazon scenarios—and consistent across cold-start ratios. If the numbers are real, that is a solid subfield-level contribution.

The genuinely new piece is the step-by-step conditional injection, which is distinct from DiffCDR’s unconditional diffusion for CDR. The six-variant study is useful; it maps the design space (where to concatenate or inject the preference signal) and shows the proposed scheme is best. The writing is clear and the method is mostly specified precisely, and using the full Amazon data rather than a filtered subset is a plus.

The soft spots are not fatal but need attention. First, there is no code, no standard deviations, and no significance tests, even though the text says results are averaged over five runs. Hyperparameters are grid-searched by MAE without a described validation split, so we cannot assess whether the margins in Table 2 reflect a stable effect. Second, Eq. (16) has a stray epsilon in the predicted posterior mean; the x0-prediction parameterization used in Eq. (17) and Eq. (21) does not contain that noise term. It is probably a typo, but it must be fixed. Third, the cold-start user’s target embedding u_i used to initialize the reverse chain (Algorithm 2) is an untrained random vector because those users have no target ratings. The w/o DM ablation in Section 4.4 uses the same random vector concatenated with h^s_i, so that ablation is not a clean control for the diffusion component. The paper never compares against a direct non-diffusion predictor mapping h^s_i to a target embedding with comparable capacity and the same L_rec objective. That matched control is needed to support the claim that diffusion, rather than the preference signal and objective, is the source of the gains.

Who this is for: people working on CDR or diffusion-based recommenders. The variant study alone is worth having in the literature.

My recommendation: send it to peer review, but require code/data release, uncertainty quantification (error bars or significance tests), a corrected Eq. (16), and the matched non-diffusion control before accepting. The core idea is worth engaging with seriously.

Referee Report

3 major / 6 minor

Summary. The manuscript proposes DMCDR, a diffusion-model-based method for cross-domain recommendation (CDR) to cold-start users. The model encodes a user's source-domain interaction history with a Transformer and average pooling into a preference signal h_i^s, corrupts the target-domain user embedding u_{i,0} with an exponential noise schedule, and learns a reverse process that takes h_i^s as a stepwise condition. Training combines a rating-reconstruction MSE over overlapping users with a diffusion objective in which the network predicts the clean target embedding; inference starts from the cold-start user's own target embedding and denoises it under classifier-free guidance. Experiments on three Amazon scenarios compare DMCDR with ten baselines and report consistent MAE and RMSE improvements, with relative MAE gains of up to 32.60% over CDRNP. Additional experiments study six diffusion variants, ablations, hyperparameters, and computational cost.

Significance. The core idea, using the reverse diffusion chain itself as the preference-transfer mechanism rather than a separate embedding-mapping function, is a reasonable and novel framing. If the reported results are robust, the paper would demonstrate a new high-performing paradigm for cold-start CDR and would extend diffusion-based recommender research to the cross-domain setting. The six-variant study is a useful attempt to isolate design choices, and the use of full Amazon data rather than subsamples is commendable. The main limitation is that the current experimental design does not yet establish that the diffusion mechanism, rather than the preference signal or the particular training objective, is responsible for the gains; a matched non-diffusion control and a clearer validation protocol are needed.

major comments (3)

[Section 3.4.2 / Algorithm 2 / Section 4.4] The inference procedure initializes the reverse chain with the cold-start user's own target-domain embedding u_i, but these users have no target-domain ratings during training, so this embedding receives no gradient and is effectively an untrained random vector. The w/o DM ablation in Section 4.4 then recommends from the concatenation of that same random vector with h_i^s, so the reported degradation reflects the inclusion of a random feature, not the removal of the denoising process. To support the central claim, please compare DMCDR against a matched non-diffusion predictor that maps h_i^s to the target representation, such as an MLP with comparable capacity trained with the same L_rec objective, and test whether initializing from h_i^s or from standard Gaussian noise changes the result.
[Section 3.3.2, Eqs. (16)-(17) and (21)] Equation (16) writes the posterior-mean predictor as sqrt(alpha_bar_{t-1}) f_theta + (...) epsilon, making mu_theta depend on the forward noise epsilon. At inference, however, epsilon is not known, and Algorithm 2 uses Eq. (21), which is the standard DDPM posterior mean in terms of f_theta and contains no epsilon. Please clarify whether Eq. (16) is only a training-time construction used to derive the x0-prediction loss in Eq. (17), and state explicitly that sampling uses Eq. (21) with the guided predictor from Eq. (20). As written, the definition of mu_theta is not the quantity evaluated in the reverse process, so the derivation is internally inconsistent.
[Section 4.1.4 / Table 2] The implementation details say that hyperparameters are grid-searched according to MAE but do not describe a validation split. If the grid search uses the held-out cold-start users whose ratings define the reported MAE, the numbers in Table 2 are optimistically selected and the comparison is not a fair test. Also, although results are averaged over five runs, no standard deviations or significance tests are reported, while Section 4.2 asserts that gains are significant. Please specify the exact split used for hyperparameter selection and report mean plus/minus standard deviation for the main comparisons, ideally with paired tests across random seeds.

minor comments (6)

[Eq. (8)] Equation (8) calls S_t a set of T uniformly spaced values but then uses S_t as a scalar; please define S_t as the t-th element of that schedule, with a concrete formula for how the values are assigned to steps.
[Section 3.2, Eq. (4)] The handling of variable-length interaction histories is not described; please specify the padding, truncation, and masking protocol used when feeding H_i^s to the Transformer layer.
[Algorithm 2] Line 4 says 'Control the strength of h_i^s' without saying how; please state that Eq. (20) is applied to f_theta before the update in Eq. (21).
[Table 2] The 18-column layout makes the table hard to read across the three scenarios; consider grouping results by dataset in separate sub-tables or using a multi-index layout.
[Figure 6] The history-length comparison reports only MAE and no error bars, so it is difficult to assess whether the differences between methods are stable across runs; please add standard deviations or confidence bands.
[General] No artifact-availability statement is provided; please indicate whether code and the exact data splits will be released to support reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the preference-guided diffusion derivation is self-contained, and the self-citation to CDRNP appears only as a baseline comparator, not as load-bearing evidence.

full rationale

The claimed derivation chain does not reduce any prediction to its own input. The preference signal h_i^s is computed from the user's source-domain interaction history via Eqs. (4)-(5), which is a distinct input from the target-domain representation being generated. The diffusion training objective in Eq. (17) is a reconstruction loss on overlapping users' target embeddings, while the recommendation loss in Eq. (10) uses target ratings; these are combined as L = L_rec + lambda*L_{t-1} in Eq. (18) rather than being defined in terms of each other. At inference, cold-start users' target ratings are withheld, so the reported MAE/RMSE are computed on data not used to fit the model. The use of u_i^{s\o} as the initial reverse-process state (Section 3.4.2) is an unusual and potentially fragile design choice, but it is not circular: the output u_i^{s\o,0} is produced by the learned denoising update in Eq. (21), not algebraically equal to the initialization or to h_i^s. Likewise, the w/o DM ablation is an experimental comparison, not a derivation step; whether it is a perfectly matched control is a correctness/experimental-design concern, not a circularity concern. The only author-overlap self-citation is the CDRNP baseline [18], which is used as a competitor and for standard evaluation protocol, not as the justification for DMCDR's mechanism. No equation in the paper collapses to its own input, no fitted parameter is renamed as a prediction, and no external uniqueness claim is imported. The paper is self-contained against external benchmarks, so the circularity score is 0.

Assumptions & free parameters 9 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard DDPM machinery, a domain transfer assumption, and a hand-designed noise schedule with several grid-searched hyperparameters. No new physical or theoretical entities are introduced; the preference guidance signal is a learned latent feature.

free parameters (9)

eta (forward noise scale) = grid search over {0.1,0.3,0.5,0.7,0.9}
Controls added noise in Eq. (8); no analytic justification given, tuned by MAE.
omega (guidance strength) = grid search over {0,1,2,3,4,5}
Controls source-preference strength in Eq. (20); tuned per scenario.
T (diffusion steps) = 200
Searched over {50,100,200,500,1000}; chosen to balance performance and computational cost.
lambda (diffusion loss weight) = 1e-2
Fixed by hand in Eq. (18).
p_uncond = 0.1
Unconditional-dropout probability in Eq. (19), set following classifier-free guidance literature.
alpha_min, alpha_max = 0.1, 10
Bounds of the custom noise schedule in Eq. (8); chosen by hand.
history length |H_i^s| = grid search over {10,20,30,40,50}
Number of source interactions encoded; affects the preference representation.
Transformer/MLP depth = grid search over {2,3,4,5,6}
Depth of the preference encoder and denoiser; tuned by MAE.
embedding dimension, batch size, learning rate = 64, 128, 0.01
Standard training choices fixed for all models; still hand-selected.

assumptions (4)

standard math Standard DDPM forward/reverse processes and the variational bound from Ho et al. apply to continuous user representations.
Section 2.1 and Eq. (3) rely on the standard DDPM formulation.
domain assumption A source-domain preference summary derived from item embeddings captures transferable user preference, and the conditional denoiser trained on overlapping users generalizes to cold-start users.
Core of the method in Sections 3.2 and 3.4; no theoretical analysis of distribution shift is provided.
ad hoc to paper Untrained target-domain embeddings of cold-start users are a usable initialization for the reverse denoising chain.
Algorithm 2 starts inference from u_i for cold-start users, even though this embedding has no target-domain training signal.
ad hoc to paper The custom exponential noise schedule in Eq. (8), with eta, alpha_min, and alpha_max, preserves personalization while still enabling denoising training.
Introduced specifically for this model in Section 3.3.1, with hyperparameters selected by grid search.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Exploring Preference-Guided Diffusion Model for Cross-Domain Recommendation." pith.science (2026). https://pith.science/paper/FVKUQ2EE

@misc{pith2026250111671,
  author       = {Pith},
  title        = {Pith review of: Exploring Preference-Guided Diffusion Model for Cross-Domain Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FVKUQ2EE}},
  note         = {Machine review of arXiv:2501.11671}
}

read the original abstract

Cross-domain recommendation (CDR) has been proven as a promising way to alleviate the cold-start issue, in which the most critical problem is how to draw an informative user representation in the target domain via the transfer of user preference existing in the source domain. Prior efforts mostly follow the embedding-and-mapping paradigm, which first integrate the preference into user representation in the source domain, and then perform a mapping function on this representation to the target domain. However, they focus on mapping features across domains, neglecting to explicitly model the preference integration process, which may lead to learning coarse user representation. Diffusion models (DMs), which contribute to more accurate user/item representations due to their explicit information injection capability, have achieved promising performance in recommendation systems. Nevertheless, these DMs-based methods cannot directly account for valuable user preference in other domains, leading to challenges in adapting to the transfer of preference for cold-start users. Consequently, the feasibility of DMs for CDR remains underexplored. To this end, we explore to utilize the explicit information injection capability of DMs for user preference integration and propose a Preference-Guided Diffusion Model for CDR to cold-start users, termed as DMCDR. Specifically, we leverage a preference encoder to establish the preference guidance signal with the user's interaction history in the source domain. Then, we explicitly inject the preference guidance signal into the user representation step by step to guide the reverse process, and ultimately generate the personalized user representation in the target domain, thus achieving the transfer of user preference across domains. Furthermore, we comprehensively explore the impact of six DMs-based variants on CDR.

Figures

Figures reproduced from arXiv: 2501.11671 by the authors.

**Figure 2.** The overall framework of DMCDR in the training phase. To achieve preference transfer, DMCDR encodes user’s inter [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** An illustration of six DMs-based variants in the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 6.** Figure 6: Effect of the length of H𝑠 𝑖 . scenario1 scenario2 scenario3 (a) GPU usage 500 1,500 2,500 3,500 GPU (MB) DMCDR CDRNP scenario1 scenario2 scenario3 (b) Parameters 0 30 60 90 par. (M) DMCDR CDRNP scenario1 scenario2 scenario3 (c) Training time 0 30 60 90 time. (s) / epo…

**Figure 7.** Figure 7: Computational cost on three CDR scenarios. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

FARM: Frequency-Aware Model for Cross-Domain Live-Streaming Recommendation
cs.IR 2025-02 conditional novelty 4.0 of 10

FARM combines discrete Fourier frequency splitting with contrastive preference alignment to improve cross-domain live-streaming recommendation at Kuaishou.

Reference graph

Works this paper leans on

58 extracted references · 20 canonical work pages · cited by 1 Pith paper

[1]

Jiangxia Cao, Xixun Lin, Xin Cong, Jing Ya, Tingwen Liu, and Bin Wang. 2022. Disencdr: Learning disentangled representations for cross-domain recommenda- tion. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval . 267–277

work page 2022
[2]

Jiangxia Cao, Jiawei Sheng, Xin Cong, Tingwen Liu, and Bin Wang. 2022. Cross- domain recommendation to cold-start users via variational information bottle- neck. In 2022 IEEE 38th International Conference on Data Engineering (ICDE). IEEE, 2209–2223

2022
[3]

Nanxin Chen, Yu Zhang, Heiga Zen, Ron J Weiss, Mohammad Norouzi, and William Chan. 2020. Wavegrad: Estimating gradients for waveform generation. arXiv preprint arXiv:2009.00713 (2020)

arXiv 2020
[4]

Zhiying Deng, Jianjun Li, Zhiqiang Guo, Wei Liu, Li Zou, and Guohui Li. 2023. Multi-view multi-aspect neural networks for next-basket recommendation. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval . 1283–1292

work page 2023
[5]

Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34 (2021), 8780–8794

2021
[6]

Chelsea Finn, Pieter Abbeel, and Sergey Levine. 2017. Model-agnostic meta- learning for fast adaptation of deep networks. In International conference on machine learning. PMLR, 1126–1135

2017
[7]

Marta Garnelo, Jonathan Schwarz, Dan Rosenbaum, Fabio Viola, Danilo J Rezende, SM Eslami, and Yee Whye Teh. 2018. Neural processes. arXiv preprint arXiv:1807.01622 (2018)

arXiv 2018
[8]

Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and LingPeng Kong. 2022. Diffuseq: Sequence to sequence text generation with diffusion models. arXiv preprint arXiv:2210.08933 (2022)

arXiv 2022

Show all 58 references

[9]

Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. In Proceedings of the 26th international conference on world wide web . 173–182

2017
[10]

Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems 33 (2020), 6840–6851. KDD ’25, August 3–7, 2025, Toronto, ON, Canada Xiaodong Li et al

2020
[11]

Jonathan Ho and Tim Salimans. 2022. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)

2022 arXiv
[12]

Guangneng Hu, Yu Zhang, and Qiang Yang. 2018. Conet: Collaborative cross networks for cross-domain recommendation. In Proceedings of the 27th ACM international conference on information and knowledge management . 667–676

2018
[13]

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
[14]

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 Knowl- edge Management. 1563–1572

2019
[15]

Diederik P Kingma and Max Welling. 2013. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 (2013)

2013 arXiv
[16]

Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. 2020. Diffwave: A versatile diffusion model for audio synthesis. arXiv preprint arXiv:2009.09761 (2020)

2020 arXiv
[17]

Pan Li and Alexander Tuzhilin. 2021. Dual metric learning for effective and efficient cross-domain recommendations. IEEE Transactions on Knowledge and Data Engineering 35, 1 (2021), 321–334

2021
[18]

Xiaodong Li, Jiawei Sheng, Jiangxia Cao, Wenyuan Zhang, Quangang Li, and Tingwen Liu. 2024. CDRNP: Cross-Domain Recommendation to Cold-Start Users via Neural Process. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining. 378–386

2024
[19]

Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. 2022. Diffusion-lm improves controllable text generation. Advances in Neural Information Processing Systems 35 (2022), 4328–4343

2022
[20]

Weiming Liu, Chaochao Chen, Xinting Liao, Mengling Hu, Yanchao Tan, Fan Wang, Xiaolin Zheng, and Yew Soon Ong. 2024. Learning Accurate and Bidirec- tional Transformation via Dynamic Embedding Transportation for Cross-Domain Recommendation. In Proceedings of the AAAI Conference...

2024
[21]

Weiming Liu, Xiaolin Zheng, Mengling Hu, and Chaochao Chen. 2022. Collab- orative filtering with attribution alignment for review-based non-overlapped cross domain recommendation. In Proceedings of the ACM web conference 2022 . 1181–1190

2022
[22]

Yuanfu Lu, Yuan Fang, and Chuan Shi. 2020. Meta-learning on heterogeneous information networks for cold-start recommendation. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining . 1563–1573

2020
[23]

Calvin Luo. 2022. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970 (2022)

2022 arXiv
[24]

Haokai Ma, Ruobing Xie, Lei Meng, Xin Chen, Xu Zhang, Leyu Lin, and Zhan- hui Kang. 2024. Plug-in diffusion model for sequential recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 38. 8886–8894

2024
[25]

Haokai Ma, Ruobing Xie, Lei Meng, Xin Chen, Xu Zhang, Leyu Lin, and Jie Zhou. 2024. Triple sequence learning for cross-domain recommendation. ACM Transactions on Information Systems 42, 4 (2024), 1–29

2024
[26]

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
[27]

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
[28]

Chang Meng, Chenhao Zhai, Yu Yang, Hengyu Zhang, and Xiu Li. 2023. Parallel knowledge enhancement based framework for multi-behavior recommendation. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management. 1797–1806

2023
[29]

Alexander Quinn Nichol and Prafulla Dhariwal. 2021. Improved denoising diffu- sion probabilistic models. In International conference on machine learning . PMLR, 8162–8171

2021
[30]

Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme
[31]

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 10684–10695

2022
[32]

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
[33]

Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli
[34]

Jiaming Song, Chenlin Meng, and Stefano Ermon. 2021. Denoising diffusion implicit models. International Conference on Learning Representations (2021)

2021
[35]

Caiqi Sun, Jiewei Gu, BinBin Hu, Xin Dong, Hai Li, Lei Cheng, and Linjian Mo
[36]

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems 30 (2017)

2017
[37]

Wenjie Wang, Yiyan Xu, Fuli Feng, Xinyu Lin, Xiangnan He, and Tat-Seng Chua
[38]

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
[39]

Ruobing Xie, Qi Liu, Liangdong Wang, Shukai Liu, Bo Zhang, and Leyu Lin
[40]

Yiyan Xu, Wenjie Wang, Fuli Feng, Yunshan Ma, Jizhi Zhang, and Xiangnan He
[41]

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
[42]

Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. 2023. Diffusion models: A comprehensive survey of methods and applications. Comput. Surveys 56, 4 (2023), 1–39

2023
[43]

Zhengyi Yang, Jiancan Wu, Zhicai Wang, Xiang Wang, Yancheng Yuan, and Xiangnan He. 2024. Generate What You Prefer: Reshaping Sequential Recommen- dation via Guided Diffusion. Advances in Neural Information Processing Systems 36 (2024)

2024
[44]

Wenhui Yu, Xiao Lin, Junfeng Ge, Wenwu Ou, and Zheng Qin. 2020. Semi- supervised collaborative filtering by text-enhanced domain adaptation. In Pro- ceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining. 2136–2144

2020
[45]

Feng Yuan, Lina Yao, and Boualem Benatallah. 2019. DARec: deep domain adaptation for cross-domain recommendation via transferring rating patterns. In Proceedings of the 28th International Joint Conference on Artificial Intelligence . 4227–4233

2019
[46]

Qian Zhang, Jie Lu, Dianshuang Wu, and Guangquan Zhang. 2018. A cross- domain recommender system with kernel-induced knowledge transfer for over- lapping entities. IEEE transactions on neural networks and learning systems 30, 7 (2018), 1998–2012

2018
[47]

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

2024 arXiv
[48]

Jiawei Zheng, Qianli Ma, Hao Gu, and Zhenjing Zheng. 2021. Multi-view denois- ing graph auto-encoders on heterogeneous information networks for cold-start recommendation. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining . 2338–2348

2021
[49]

Feng Zhu, Yan Wang, Chaochao Chen, Guanfeng Liu, Mehmet Orgun, and Jia Wu
[50]

Yongchun Zhu, Kaikai Ge, Fuzhen Zhuang, Ruobing Xie, Dongbo Xi, Xu Zhang, Leyu Lin, and Qing He. 2021. Transfer-meta framework for cross-domain recom- mendation to cold-start users. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in In...

2021
[51]

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 . 1507–1515

2022
[53]

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
[2012]

arXiv preprint arXiv:1205.2618 (2012)

BPR: Bayesian personalized ranking from implicit feedback. arXiv preprint arXiv:1205.2618 (2012)

2012 arXiv
[2015]

In International conference on machine learning

Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning . PMLR, 2256–2265
[2018]

In Proceedings of the 27th International Joint Conference on Artificial Intelligence

A deep framework for cross-domain and cross-system recommendations. In Proceedings of the 27th International Joint Conference on Artificial Intelligence
[2022]

In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining

Contrastive cross-domain recommendation in matching. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining . 4226–4236
[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
[2024]

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

Diffusion Models for Generative Outfit Recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1350–1359

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Jiangxia Cao, Xixun Lin, Xin Cong, Jing Ya, Tingwen Liu, and Bin Wang. 2022. Disencdr: Learning disentangled representations for cross-domain recommenda- tion. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval . 267–277

work page 2022

[2] [2]

Jiangxia Cao, Jiawei Sheng, Xin Cong, Tingwen Liu, and Bin Wang. 2022. Cross- domain recommendation to cold-start users via variational information bottle- neck. In 2022 IEEE 38th International Conference on Data Engineering (ICDE). IEEE, 2209–2223

2022

[3] [3]

Nanxin Chen, Yu Zhang, Heiga Zen, Ron J Weiss, Mohammad Norouzi, and William Chan. 2020. Wavegrad: Estimating gradients for waveform generation. arXiv preprint arXiv:2009.00713 (2020)

arXiv 2020

[4] [4]

Zhiying Deng, Jianjun Li, Zhiqiang Guo, Wei Liu, Li Zou, and Guohui Li. 2023. Multi-view multi-aspect neural networks for next-basket recommendation. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval . 1283–1292

work page 2023

[5] [5]

Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34 (2021), 8780–8794

2021

[6] [6]

Chelsea Finn, Pieter Abbeel, and Sergey Levine. 2017. Model-agnostic meta- learning for fast adaptation of deep networks. In International conference on machine learning. PMLR, 1126–1135

2017

[7] [7]

Marta Garnelo, Jonathan Schwarz, Dan Rosenbaum, Fabio Viola, Danilo J Rezende, SM Eslami, and Yee Whye Teh. 2018. Neural processes. arXiv preprint arXiv:1807.01622 (2018)

arXiv 2018

[8] [8]

Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and LingPeng Kong. 2022. Diffuseq: Sequence to sequence text generation with diffusion models. arXiv preprint arXiv:2210.08933 (2022)

arXiv 2022