Pith. sign in

REVIEW 2 major objections 8 minor 1 cited by

Beyond Negative Transfer: Disentangled Preference-Guided Diffusion for Cross-Domain Sequential Recommendation

T0 review · 2 major / 8 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A diffusion model that separates shared and domain-specific user preferences outperforms existing cross-domain sequential recommenders.

desk verdict Plausible diffusion-based CDSR model with a loosely constrained disentanglement mechanism, undermined by internal numerical contradictions that make its SOTA claims unverifiable. read the letter →

arxiv 2509.00389 v1 pith:B3ZWRLHE submitted 2025-08-30 cs.IR cs.AIcs.SI

Xiaoxin Ye , Chengkai Huang , Hongtao Huang , Lina Yao This is my paper · ORCID
classification cs.IRcs.AIcs.SI
keywords Cross-DomainSequentialRecommendationDiffusionModelsDisentangledRepresentationLearningNegativeTransferPreference-GuidedDenoisingContrastive
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 claims that cross-domain sequential recommendation fails when models merge all user behaviors into one sequence, because shared interests, domain-specific tastes, and accidental clicks become entangled and drag each other down. DPG-Diff tackles this by decomposing user preferences into domain-invariant and domain-specific parts that jointly guide a diffusion denoising process, so the model reconstructs a clean user representation while suppressing conflicting or noisy signals. On two real-world Amazon cross-domain benchmarks, the paper reports consistent improvements over existing baselines, with the largest gains on Movie/Book. If true, this makes diffusion models a viable generative paradigm for cross-domain sequential recommendation.

What carries the argument

The machinery is a three-part architecture: a Disentangled Encoder with two independent self-attention encoders (one per domain) that produce domain-specific representations; a Disentangled Preference-Guided Denoiser that uses cross-attention to condition the reverse diffusion process on a fused guidance vector combining both domain signals; and a tri-view contrastive loss that aligns the denoised embedding, the fused embedding, and an augmented sequence embedding. The claimed work of the machinery is to separate domain-invariant preference from domain-specific preference and noise, so that the guidance vector carries transferable signal while the denoiser filters out conflicting and irrelev

What would settle it

Train a variant in which the guidance vector is the plain concatenation of the two domain encodings without any learned fusion, or a variant with a single shared encoder for both domains, and compare metrics on the same datasets. If performance is statistically indistinguishable from DPG-Diff, the disentangled-preference guidance is not the active ingredient behind the reported gains.

Watch

Extended reading notes

Core claim

The central claim is that a diffusion model whose reverse denoising is guided by disentangled preferences outperforms both sequential recommenders and existing cross-domain sequential recommenders. The model, DPG-Diff, first encodes each domain's sequence with its own transformer encoder to obtain domain-specific representations, fuses them into a cross-domain guidance vector, then trains a denoiser that reconstructs the user's item embedding from corrupted versions using cross-attention conditioned on that guidance. A tri-view contrastive loss aligns the denoised representation, the fused guidance, and an augmented view. The paper reports that this setup, trained with a cross-entropy recomm

Load-bearing premise

The load-bearing premise is that two independent self-attention encoders, one per domain, followed by a simple fusion, actually separate domain-invariant from domain-specific preferences; no explicit disentanglement loss or orthogonality constraint is imposed, so if the encoders simply learn domain-specific encodings that are concatenated, the claimed mechanism is unsupported.

Editorial extensions

If this is right

  • If DPG-Diff's claims hold, diffusion-based generative models become a competitive paradigm for cross-domain sequential recommendation, not just single-domain SR or static cross-domain recommendation.
  • Separating guidance into domain-invariant and domain-specific components before fusion can reduce negative transfer in cross-domain sequential settings.
  • The method's robustness under injected noise suggests that guided denoising can act as an active noise filter for sequential user behavior.
  • The unified scoring function (cross-domain denoised representation plus domain-specific signal) enables prediction in both domains from a single generative pass.
  • Large reported gains on Movie/Book indicate existing CDSR models still leave substantial headroom on heterogeneous domain pairs.

Reading between the lines

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

  • The paper does not impose an explicit disentanglement loss or orthogonality constraint, so whether the two encoders actually separate shared from specific preferences is an architectural claim, not a verified property; a reader should treat that as an open empirical question.
  • The reported gains could partly come from the extra parameters and contrastive objective rather than from the guidance mechanism; an ablation that replaces the fused guidance with a plain concatenation or a single shared encoder would isolate the active ingredient.
  • The noise-injection robustness study uses synthetic noise; testing on naturally noisy interactions (e.g., misclicks inferred from dwell time or return behavior) would show whether the denoiser filters real-world noise, not just injected randomness.
  • Extending the same preference-guided denoising to three or more domains, which the paper lists as future work, is a natural test of whether the disentanglement scales beyond pairwise domain setups.
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

2 major / 8 minor

Summary. The manuscript proposes DPG-Diff, a diffusion-based model for Cross-Domain Sequential Recommendation (CDSR). The model encodes source- and target-domain user sequences with two independent self-attention encoders, fuses the outputs into a guidance representation, and uses a cross-attention-conditioned denoiser to reconstruct user embeddings. A tri-view contrastive loss aligns the denoised, fused, and augmented views. The paper reports evaluations on Amazon Movie-Book and Food-Kitchen, claiming consistent state-of-the-art improvements, plus ablation, robustness, and efficiency analyses. The proposed integration is plausible, but the experimental verification is internally inconsistent and the disentanglement mechanism is not explicitly enforced.

Significance. If the results were reliable, DPG-Diff would be a meaningful contribution: it is, to the authors' knowledge, the first diffusion-based model designed for CDSR, and the idea of using disentangled preference guidance to avoid negative transfer and to denoise sequential behavior is appealing. The loss design is conventional and the framework is reusable. However, the current manuscript does not support its central empirical claim because (i) Table 3 contradicts Table 1, (ii) the ablation text reverses a result, and (iii) no code, data, or standard deviations are provided. The paper also does not formally enforce the 'disentanglement' that motivates the method. These issues are fixable in principle, but the evidence as presented is not sufficient.

major comments (2)
  1. [Ablation Study (RQ2), Tables 1 and 3] The full-model NDCG@10 values in Table 3 are 13.50/5.58/13.50/5.58 for Food/Kitchen/Movie/Book, whereas Table 1 reports 13.07/7.01/13.34/5.79 for the same model. The Table 3 row duplicates values across datasets (Food=Movie, Kitchen=Book) and disagrees with Table 1 on every dataset. In addition, the text states that adding the Disentangled Encoder 'improves NDCG@10 from 5.95 to 5.83' on Movie, but 5.83 < 5.95, i.e., a degradation. The paper's central claim is the consistent SOTA improvement in Table 1, supported by a t-test; however, no standard deviations or run counts are reported. Without code or data, an external check is impossible, and the two tables cannot both be correct. Please correct the tables, reconcile the ablation narrative, and provide reproducible experimental details.
  2. [Disentangled Encoder, Eq. (8)] The method is described as disentangling domain-invariant, domain-specific, and noise signals, but the architecture and losses do not enforce such a decomposition. The two encoders are independent self-attention/MLP stacks, and no orthogonality, independence, or mutual-information constraint separates gx and gy or their shared components. The fusion into gd is not specified by an equation, and Lrec and Ltri-cl only supervise classification and alignment. Thus the performance gains cannot be attributed to 'disentanglement' rather than to having two domain encoders plus a fusion. Please either add an explicit disentanglement objective or provide probing/ablation evidence that the learned representations indeed separate the three signals.
minor comments (8)
  1. [Table 3 caption] The table header does not state the metric; the text indicates NDCG@10, but the caption should say so explicitly.
  2. [Eq. (15)] The equation has mismatched parentheses; it should presumably read \hat{p}_y = softmax((\hat{x}_0 + \hat{g}_y)^\top E_y).
  3. [Algorithm 1, line 9] Line 9 says 'Compute diffusion loss: Lrec', but Lrec is the recommendation loss defined in Eq. (10).
  4. [Inference paragraph] The text refers to 'Diff-Rec' instead of DPG-Diff when describing the inference procedure.
  5. [Robustness Study (RQ3), Figure 3] The robustness comparison includes only one baseline (DREAM) and no variance or error bars, making the claimed robustness advantage hard to quantify.
  6. [Model Efficiency Analysis (RQ4), Figure 4] Figure 4 plots NDCG versus inference steps, not wall-clock time or memory usage, so RQ4 (computational cost) is not directly answered.
  7. [Supplementary material] The text states 'More details are available in the supplementary material', but no supplement is included with the arXiv submission. Please provide it or remove the reference.
  8. [Typos] There are several typos, e.g., 'to or best knowledge' (Abstract), 'contrative' (Introduction), 'donates' instead of 'denotes' (Eq. (10) discussion), and 'domain-specific components and domain-specific components' (Methodology).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the empirical SOTA claim is self-contained; table inconsistencies are a correctness, not circularity, issue.

full rationale

The paper's central claim is an empirical performance comparison (Table 1) against external baselines on standard Amazon benchmarks, following the protocol of C2DREIF. The model is trained with conventional losses (diffusion MSE, cross-entropy recommendation loss, contrastive loss) and evaluated on held-out interactions. There is no parameter fitted to a subset and then reported as a prediction of a closely related quantity, no result that reduces to an equation's definition, and no load-bearing self-citation chain: the cited prior works (DiffuRec, DMCDR, CDCDR, DREAM, C2DREIF) are baselines or background, and the 'first diffusion-based CDSR' claim is a positioning statement, not an imported uniqueness theorem. The 'disentanglement' terminology is not enforced by an explicit loss, making the mechanism interpretation under-specified, but that is an evidentiary/interpretation concern, not circularity: the components are not defined in terms of the outcome they are claimed to explain. The ablation narrative contains internal numerical inconsistencies (e.g., text says adding the Disentangled Encoder 'improves NDCG@10 from 5.95 to 5.83' on Movie although 5.83 < 5.95, and Table 3's full-model NDCG@10 values differ from Table 1 and are duplicated across domains), which undermine reliability of the reported gains; however, inconsistency between two tables is not a reduction of a derivation to its inputs. Therefore no circular step is exhibited.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The model does not introduce new physical or conceptual entities. It uses existing diffusion and Transformer machinery with a new way to combine them. The main scientific risk is not a new entity but an unverified assumption about disentanglement.

free parameters (5)
  • Embedding size = 512
    Reported in Implementation Details; chosen by hand and likely tuned on validation.
  • Training epochs = 100
    Reported; expected to affect convergence and final performance.
  • Diffusion timesteps T
    Not reported; standard hyperparameter for diffusion models, affects training and inference.
  • Noise schedule beta_t
    Not reported; standard in DDPM, but the specific schedule is a choice.
  • Batch size
    Not reported; can affect optimization and results.
assumptions (4)
  • domain assumption Gaussian diffusion can model user behavior sequences as noisy data and denoised embeddings as preferences.
    The entire approach relies on the assumption that the forward corruption and reverse denoising over item embeddings is a meaningful generative model of user behavior. Invoked throughout the Methodology.
  • domain assumption Independent encoders per domain naturally separate domain-invariant from domain-specific preferences.
    No explicit disentanglement loss is used; the paper assumes that the architecture itself achieves this separation. Invoked in the Disentangled Encoder and the guidance fusion.
  • domain assumption Cross-domain knowledge transfer improves recommendation performance in the two-domain setting.
    The problem setup assumes that borrowing information between domains is beneficial, which motivated all CDSR baselines. Standard in the field.
  • domain assumption Sampled negative items (999 per test case) yield unbiased evaluation.
    Standard protocol for large-item recommender evaluation, but it relies on the sample being representative. Invoked in the Experiment setup.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Negative Transfer: Disentangled Preference-Guided Diffusion for Cross-Domain Sequential Recommendation." pith.science (2026). https://pith.science/paper/B3ZWRLHE

@misc{pith2026250900389,
  author       = {Pith},
  title        = {Pith review of: Beyond Negative Transfer: Disentangled Preference-Guided Diffusion for Cross-Domain Sequential Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B3ZWRLHE}},
  note         = {Machine review of arXiv:2509.00389}
}
read the original abstract

Cross-Domain Sequential Recommendation (CDSR) leverages user behaviors across domains to enhance recommendation quality. However, naive aggregation of sequential signals can introduce conflicting domain-specific preferences, leading to negative transfer. While Sequential Recommendation (SR) already suffers from noisy behaviors such as misclicks and impulsive actions, CDSR further amplifies this issue due to domain heterogeneity arising from diverse item types and user intents. The core challenge is disentangling three intertwined signals: domain-invariant preferences, domain-specific preferences, and noise. Diffusion Models (DMs) offer a generative denoising framework well-suited for disentangling complex user preferences and enhancing robustness to noise. Their iterative refinement process enables gradual denoising, making them effective at capturing subtle preference signals. However, existing applications in recommendation face notable limitations: sequential DMs often conflate shared and domain-specific preferences, while cross-domain collaborative filtering DMs neglect temporal dynamics, limiting their ability to model evolving user preferences. To bridge these gaps, we propose \textbf{DPG-Diff}, a novel Disentangled Preference-Guided Diffusion Model, the first diffusion-based approach tailored for CDSR, to or best knowledge. DPG-Diff decomposes user preferences into domain-invariant and domain-specific components, which jointly guide the reverse diffusion process. This disentangled guidance enables robust cross-domain knowledge transfer, mitigates negative transfer, and filters sequential noise. Extensive experiments on real-world datasets demonstrate that DPG-Diff consistently outperforms state-of-the-art baselines across multiple metrics.

Figures

Figures reproduced from arXiv: 2509.00389 by the authors.

Figure 1
Figure 1. The complexity of user behavior sequences in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. DPG-Diff integrates a Disentangled Encoder, a Disentangled Preference Guided Denoiser (DPG Denoiser) and Tri￾view CL within a diffusion recommendation architecture. The input is a cross-domain sequence Sc, composed of Sx for domain x, and Sy for domain y. Sc may contains three disentangled preference: domain-invariant, conflicting, and ad-hoc noise. The Disentangled Encoder extracts representations from each domain,… view at source ↗
Figure 3
Figure 3. Robustness comparison between DPG-Diff and [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

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

  1. Beyond Item Order: Temporal Gap Tokenization for Generative Recommendation with Semantic IDs

    cs.IR 2026-07 conditional novelty 5.0 of 10

    Interleaving fixed log-scale gap tokens with semantic IDs, plus TA-FAMAE temporal regularization, consistently beats ReSID and other SID generative baselines on Amazon sequential recommendation.

Reference graph

Works this paper leans on

37 extracted references · 31 canonical work pages · cited by 1 Pith paper

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Bian, Q.; de Carvalho, M.; Li, T.; Xu, J.; Fang, H.; and Ke, Y. 2025. ABXI: Invariant Interest Adaptation for Task-Guided Cross-Domain Sequential Recommendation. In Proceedings of the ACM on Web Conference 2025, WWW '25, 3183–3192. New York, NY, USA: Association for Computing Machinery. ISBN 9798400712746

    work page 2025

  4. [4]

    Cao, J.; Cong, X.; Sheng, J.; Liu, T.; and Wang, B. 2022. Contrastive Cross-Domain Sequential Recommendation. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, CIKM '22, 138–147. New York, NY, USA: Association for Computing Machinery. ISBN 9781450392365

    work page 2022

  5. [5]

    Dang, Y.; Yang, E.; Guo, G.; Jiang, L.; Wang, X.; Xu, X.; Sun, Q.; and Liu, H. 2024. TiCoSeRec: Augmenting Data to Uniform Sequences by Time Intervals for Effective Recommendation. IEEE Transactions on Knowledge and Data Engineering, 36(6): 2686--2700

    work page 2024

  6. [6]

    Dhariwal, P.; and Nichol, A. 2021 a . Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34: 8780--8794

    work page 2021

  7. [7]

    Dhariwal, P.; and Nichol, A. Q. 2021 b . Diffusion Models Beat GANs on Image Synthesis. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, 8780--8794

    work page 2021

  8. [8]

    Ho, J.; Jain, A.; and Abbeel, P. 2020. Denoising Diffusion Probabilistic Models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual

    work page 2020

Show all 37 references
  1. [9]

    Huang, C.; Huang, H.; Yu, T.; Xie, K.; Wu, J.; Zhang, S.; Mcauley, J.; Jannach, D.; and Yao, L. 2025. A Survey of Foundation Model-Powered Recommender Systems: From Feature-Based, Generative to Agentic Paradigms. arXiv preprint arXiv:2504.16420

    2025 arXiv

  2. [10]

    Huang, C.; Wang, S.; Wang, X.; and Yao, L. 2023 a . Dual contrastive transformer for hierarchical preference modeling in sequential recommendation. In Proceedings of the 46th international acm sigir conference on research and development in information retrieval, 99--109

    2023

  3. [11]

    Huang, C.; Wang, S.; Wang, X.; and Yao, L. 2023 b . Modeling temporal positive and negative excitation for sequential recommendation. In Proceedings of the ACM Web Conference 2023, 1252--1263

    2023

  4. [12]

    Huang, C.; Yu, T.; Xie, K.; Zhang, S.; Yao, L.; and McAuley, J. 2024 a . Foundation models for recommender systems: A survey and new perspectives. arXiv preprint arXiv:2402.11143

    2024 arXiv

  5. [13]

    Huang, H.; Huang, C.; Yu, T.; Chang, X.; Hu, W.; McAuley, J.; and Yao, L. 2024 b . Dual conditional diffusion models for sequential recommendation. arXiv preprint arXiv:2410.21967

    2024 arXiv

  6. [14]

    Kang, W.-C.; and McAuley, J. 2018. Self-Attentive Sequential Recommendation . In 2018 IEEE International Conference on Data Mining (ICDM), 197--206. Los Alamitos, CA, USA: IEEE Computer Society

    2018

  7. [15]

    P.; and Ba, J

    Kingma, D. P.; and Ba, J. 2015. Adam: A Method for Stochastic Optimization. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings

    2015

  8. [16]

    P.; and Welling, M

    Kingma, D. P.; and Welling, M. 2013. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114

    2013 arXiv

  9. [17]

    Krichene, W.; and Rendle, S. 2020. On Sampled Metrics for Item Recommendation. In KDD 2020

    2020

  10. [18]

    Li, C.; Zhao, M.; Zhang, H.; Yu, C.; Cheng, L.; Shu, G.; Kong, B.; and Niu, D. 2022. RecGURU : Adversarial Learning of Generalized User Representations for Cross-Domain Recommendation. In Proceedings of the Fifteenth ACM International Conference on Web Search and Data Mining . ACM

    2022

  11. [19]

    Li, H.; Li, J.; Ma, W.; Sun, P.; Wu, H.; Wang, J.; Yang, Y.; Zhang, M.; and Ma, S. 2025 a . CD-CDR: Conditional Diffusion-based Item Generation for Cross-Domain Recommendation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Informa...

    2025

  12. [20]

    Li, X.; Tang, H.; Sheng, J.; Zhang, X.; Gao, L.; Cheng, S.; Yin, D.; and Liu, T. 2025 b . Exploring Preference-Guided Diffusion Model for Cross-Domain Recommendation. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.1, KDD '25, 719–728....

    2025

  13. [21]

    Li, Z.; Sun, A.; and Li, C. 2024. DiffuRec: A Diffusion Model for Sequential Recommendation. ACM Trans. Inf. Syst. , 42(3): 66:1--66:28

    2024

  14. [22]

    Ma, M.; Ren, P.; Lin, Y.; Chen, Z.; Ma, J.; and Rijke, M. d. 2019. -Net: A Parallel Information-Sharing Network for Shared-Account Cross-Domain Sequential Recommendations. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information ...

    2019

  15. [23]

    Ni, R.; Cai, W.; and Jiang, Y. 2024. Contrastive cross-domain sequential recommendation via emphasized intention features. Neural Networks, 179: 106488

    2024

  16. [24]

    Pan, L.; Pan, W.; Wei, M.; Yin, H.; and Ming, Z. 2025. A Survey on Sequential Recommendation. arXiv:2412.12770

    2025 arXiv

  17. [25]

    Park, C.; Kim, T.; Choi, T.; Hong, J.; Yu, Y.; Cho, M.; Lee, K.; Ryu, S.; Yoon, H.; Choi, M.; and Choo, J. 2023. Cracking the Code of Negative Transfer: A Cooperative Game Theoretic Approach for Cross-Domain Sequential Recommendation. In Proceedings of the 32nd ACM Internation...

    2023

  18. [26]

    Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z.; Gimelshein, N.; Antiga, L.; et al. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32

    2019

  19. [27]

    Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022. High-Resolution Image Synthesis with Latent Diffusion Models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022 , 10674--10685. IEEE

    2022

  20. [28]

    Sun, W.; Ma, M.; Ren, P.; Lin, Y.; Chen, Z.; Ren, Z.; Ma, J.; and de Rijke, M. 2023. Parallel Split-Join Networks for Shared Account Cross-Domain Sequential Recommendations. IEEE Trans. on Knowl. and Data Eng., 35(4): 4106–4123

    2023

  21. [29]

    Wang, X.; Yue, H.; Wang, Z.; Xu, L.; and Zhang, J. 2023. Unbiased and Robust: External Attention-enhanced Graph Contrastive Learning for Cross-domain Sequential Recommendation . In 2023 IEEE International Conference on Data Mining Workshops (ICDMW), 1526--1534. Los Alamitos, C...

    2023

  22. [30]

    Wang, Y.; Xie, Q.; Bao, Z.; Tang, M.; Li, L.; and Liu, Y. 2025. Enhancing Transferability and Consistency in Cross-Domain Recommendations via Supervised Disentanglement

    2025

  23. [31]

    Xiao, S.; Chen, R.; Han, Q.; Lai, R.; Song, H.; and Li, L. 2023. Proxy-Aware Cross-Domain Sequential Recommendation. In 2023 International Joint Conference on Neural Networks (IJCNN), 1--8

    2023

  24. [32]

    Yang, L.; Zhang, Z.; Song, Y.; Hong, S.; Xu, R.; Zhao, Y.; Zhang, W.; Cui, B.; and Yang, M. 2024. Diffusion Models: A Comprehensive Survey of Methods and Applications. ACM Comput. Surv. , 56(4): 105:1--105:39

    2024

  25. [33]

    Yang, Z.; Wu, J.; Wang, Z.; Wang, X.; Yuan, Y.; and He, X. 2023. Generate What You Prefer: Reshaping Sequential Recommendation via Guided Diffusion. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIP...

    2023

  26. [34]

    Ye, X.; Li, Y.; and Yao, L. 2023. DREAM: Decoupled Representation via Extraction Attention Module and Supervised Contrastive Learning for Cross-Domain Sequential Recommender. In Proceedings of the 17th ACM Conference on Recommender Systems, RecSys '23, 479–490. New York, NY, U...

    2023

  27. [35]

    Zhang, D.; Liu, Z.; Jia, W.; Wu, F.; Liu, H.; and Tan, J. 2024. Dual Attention Graph Convolutional Network for Relation Extraction. IEEE Transactions on Knowledge and Data Engineering, 36(2): 530--543

    2024

  28. [36]

    Zhang, J.; Duan, H.; Guo, L.; Xu, L.; and Wang, X. 2023. Towards Lightweight Cross-Domain Sequential Recommendation via External Attention-Enhanced Graph Convolution Network. In Database Systems for Advanced Applications: 28th International Conference, DASFAA 2023, Tianjin, Ch...

    2023

  29. [37]

    Zhou, G.; Huang, C.; Chen, X.; Xu, X.; Wang, C.; Zhu, L.; and Yao, L. 2023. Contrastive counterfactual learning for causality-aware interpretable recommender systems. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, 3564--3573

    2023

Pith tools

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