Pith. sign in

REVIEW 3 major objections 5 minor 48 references

Enhancing Transferability and Consistency in Cross-Domain Recommendations via Supervised Disentanglement

T0 review · 3 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 A GNN-enhanced encoder–decoder with anchor-based supervised disentanglement improves cross-domain recommendation accuracy by up to 11.59% over the best baseline.

desk verdict A solid, incremental cross-domain recommendation paper with plausible empirical gains, but the 'supervised disentanglement' mechanism is under-analyzed—the decoder loss may be absorbed by auxiliary mapping networks rather than constraining the encoder. read the letter →

arxiv 2507.17112 v1 pith:VMNZFPDN submitted 2025-07-23 cs.IR

Yuhan Wang , Qing Xie , Zhifeng Bao , Mengzi Tang , Lin Li , Yongjian Liu This is my paper
classification cs.IR
keywords cross-domainrecommendationdisentangledrepresentationlearninggraphconvolutionalnetworkscontrastivesuperviseddisentanglementanchor-basedlossknowledgetransfersystems
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 the two obstacles to disentanglement-based cross-domain recommendation—separating features before collaborative signals are extracted, and using unsupervised disentanglement objectives that lack task guidance—can both be overcome by a single encoder–decoder framework. Its method, DGCDR, first runs a GNN over each domain's interaction graph to obtain enriched embeddings, then disentangles those embeddings into domain-shared and domain-specific spaces under orthogonality and cross-domain similarity constraints, and finally applies an anchor-based hierarchical contrastive loss that explicitly supervises the ordering of shared, GNN-enhanced, and specific features. On six cross-domain tasks built from Amazon and Douban data, the paper reports state-of-the-art results with improvements up to 11.59% over the best baseline, together with t-SNE and attention analyses suggesting that the learned separation is meaningful and transferable. If the claim is right, task-specific supervision is a practical substitute for hand-labeled disentanglement signals in cross-domain recommendation.

What carries the argument

The anchor-based hierarchical contrastive decoder loss. Using the GNN-enhanced user embedding of one domain as an anchor, the decoder compares the similarities of cross-domain transformed shared ($\hat{\mathbf{e}}_u^{c,A}$), GNN-enhanced ($\hat{\mathbf{e}}_u^{g,A}$), and specific ($\hat{\mathbf{e}}_u^{s,A}$) features to the anchor, and enforces the ranking shared $>$ GNN $>$ specific through a ratio-based pairwise InfoNCE-style loss (Eq.\ 9). This single mechanism simultaneously supervises cross-domain alignment (shared features pulled toward the anchor) and intra-domain disentanglement (specific features pushed away), replacing hand-labeled supervision with a task-derived ordering.

What would settle it

Train DGCDR with the hierarchy in Eq. (9) reversed (domain-specific features closest to the anchor, shared furthest); if performance does not drop materially on the same six tasks, the claimed benefit of the supervised ordering is not causal. Alternatively, directly measure dot-product similarities between the anchor and the three transformed feature types on held-out data: if domain-specific features are on average closer to the anchor than GNN-enhanced features are, the learned representations violate the assumed hierarchy.

Watch

Extended reading notes

Core claim

The paper's central discovery is that explicit, decoder-based supervision of the disentanglement process—rather than more contrastive regularization or generative priors—is what makes disentangled features both consistent within a domain and aligned across domains. Concretely, DGCDR establishes a hierarchical preference among three feature types with respect to a GNN-enhanced anchor embedding: transformed domain-shared features must be most similar to the anchor, GNN-enhanced features second, and transformed domain-specific features least similar. The ratio-based pairwise contrastive losses that encode this ordering provide the 'supervised' signal that stabilizes disentanglement and yields the reported gains.

Load-bearing premise

The decoder loss presumes the ranking that domain-shared features should be more similar to the GNN anchor than GNN-enhanced features, which in turn should be more similar than domain-specific features; if that ordering is wrong, the supervision actively misaligns features.

Editorial extensions

If this is right

  • Disentanglement for cross-domain recommendation can be supervised without annotations: the hierarchical ordering relative to a GNN anchor supplies the task guidance.
  • Pre-separation is unnecessary; extracting collaborative signals first and then disentangling preserves interaction information and improves accuracy, as the full-model results against GNN-contrastive baselines indicate.
  • Personalized attention-based fusion of shared and specific features is a major contributor to the gains, since ablations show removing it causes the largest performance drop (22.69% on average).
  • The same architecture transfers across domain pairs with different relatedness, achieving its largest relative gain (11.59%) on the weakly correlated Elec&Cloth pair and 10.53% on the more correlated Sport&Cloth pair.
  • The ratio-based contrastive form of the hierarchical loss is preferred over a margin-based form, which underperforms and converges more slowly in the authors' experiments.

Reading between the lines

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

  • If the hierarchical ordering assumption holds generally, it offers a cheap supervision signal that could be ported to other disentanglement tasks, such as multi-domain user modeling or cold-start personalization, where explicit labels are unavailable.
  • The attention analysis suggests a testable design rule: in weakly related domains the model should weight shared features more, while in strongly related domains it should weight specific features more; future work could exploit this as a prior or regularizer rather than learning attention from scratch.
  • The reported hyperparameter sensitivity (optimal encoder/decoder loss weights flip across datasets) indicates that the balance between the two supervision signals is data-dependent, so an adaptive weighting strategy would likely be needed before the method deploys in dynamic environments.
  • The experiments use a full user-overlap setting; the authors note the method also works for partial overlap, but a systematic study on partial-overlap data would clarify how much of the benefit depends on having the same users in both domains.
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

3 major / 5 minor

Summary. The paper proposes DGCDR, a cross-domain recommendation method that first applies GCN to enrich user/item embeddings, then disentangles them into domain-shared and domain-specific features through an encoder, and finally applies an anchor-based hierarchical contrastive decoder loss to supervise cross-domain alignment. The model is evaluated on six tasks built from Amazon and Douban domain pairs and compared against single-domain and cross-domain baselines. The paper reports accuracy gains on most metrics, an ablation study, t-SNE visualizations, attention distribution analysis, and hyperparameter analysis, and it releases code on GitHub.

Significance. If the attribution is sound, the paper makes a useful applied contribution: it tackles the known instability of unsupervised disentanglement in cross-domain recommendation by injecting task-specific supervision, and its GNN-before-disentanglement ordering is a sensible response to the pre-separation limitation of prior GNN-contrastive methods. The experimental protocol is standard: held-out target-domain splits, multiple real-world datasets, paired significance tests, ablations, and public code. The main reservation is that the decoder loss is computed on MLP-gated transformed features rather than on the encoder outputs directly, so the current evidence does not establish that the proposed 'supervised disentanglement' actually constrains the encoder's representations.

major comments (3)
  1. [Sec. 3.3, Eqs. (7)-(10)] The central claim that the decoder loss L_de provides supervised disentanglement is not supported by the present analysis. Equations (7) and (8) transform each feature by element-wise multiplication with a sigmoid-gated MLP output, and Eq. (9) measures dot-product similarity to the anchor. Because every gate value lies in (0,1), the mapping networks alone can satisfy the hierarchical ranking: setting gates near 0 for e_s,B drives f(anchor, hat e_s,A) toward 0, while keeping gates near 1 for e_c,B inflates the positive similarity. Since these gated features are used only in L_de and not in the prediction path of Eq. (5), the decoder objective can be minimized without enforcing any separation in the encoder itself. The -Dec ablation and the t-SNE analysis in Sec. 4.5 do not isolate raw encoder outputs from the gated transformed versions; without a control (e.g., fixed or identity mapping network) or a gradient analysis showing that L_de changes e_c and e_s, Eq. (9) does not substantiate the claimed supervised disentanglement.
  2. [Sec. 4.3, Table 3] The claim that DGCDR 'consistently outperforms baselines across almost all metrics' is overstated. On the Elec row, DGCDR's MRR (0.0353) is below DRLCDR's (0.0354), and on the second Cloth row, DGCDR's MRR (0.0176) is below both DRLCDR (0.0200) and BiTGCF (0.0168). Moreover, no standard deviations are reported for any cell; the only significance evidence is the paired t-test asterisks. Several reported gains are under 1% relative (for example, first Cloth MRR 0.0278 vs 0.0276), so without variance information the practical significance and run-to-run stability of the claimed improvements cannot be assessed. Please report mean and standard deviation across the five runs and revise the 'consistently outperforms' wording accordingly.
  3. [Sec. 3.3, Eq. (9)] The supervision in the decoder rests entirely on the hierarchical ordering that domain-shared features should be more similar to the anchor than GNN-enhanced features, which in turn should be more similar than domain-specific features. This ordering is introduced as a designer-chosen inductive bias, but the paper provides no empirical or information-theoretic justification for it. If the ordering is reversed for some user-item pairs, L_de will actively misalign features rather than disentangle them. Because this ranking is the only source of 'supervised' signal in the decoder, it should be validated directly, for instance by comparing against reversed or permuted orderings or by measuring whether the learned representation order agrees with held-out domain labels.
minor comments (5)
  1. [Sec. 3.2, Eq. (3)] The encoder loss notation adds the parameter set Theta_enc to the scalar loss value; this is not a loss term and should be removed or replaced with an explicit L2 regularizer in Eq. (12).
  2. [Table 4] The column header 'DRCDR†' appears to be a typo for 'DRLCDR†'; please correct it for consistency with the rest of the paper.
  3. [Sec. 4.1, Table 2] The per-row statistics for AmazonCloth and AmazonSport/Cloth are hard to interpret because the table appears to list one row per domain yet the #OverlapUsers values differ between rows of the same domain pair; please clarify how each row maps to a domain and why the overlap counts differ.
  4. [Sec. 4.4] The ablation paragraph states that MarDec 'generally requires more epochs to converge' but gives no numbers; please quantify the convergence behavior, for example with epochs-to-best or a convergence curve.
  5. [Sec. 3.3] The relationship between the proposed decoder loss and the modality-disentangled loss of Han et al. [9] should be stated more precisely; currently the paper says only that inspiration was drawn from that work, leaving the exact similarities and differences unclear.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant circularity: the headline SOTA claim is held-out empirical; one minor self-referential component in the decoder's 'supervised' disentanglement.

  1. self definitional [Sec. 3.3, Eqs. (7)-(10); Sec. 4.5, Fig. 2]
    "we use GNN-enhanced representations e_g,A and e_g,B as anchors, and compare their normalized similarity scores with those of the transformed domain-shared, domain-specific, and GNN-enhanced features from the other domain. This formulation follows the InfoNCE loss framework [25], where each anchor is encouraged to be more similar to its positive sample than to the negative one, thereby establishing a supervised hierarchical ordering of representations."

    The claimed 'explicit supervisory signal' is not an external label: the Eq. (9) ranking compares the anchor e_g with transformed features that are deterministic functions of the model's own encoder outputs (Eq. 7-8: e_hat = e ⊙ sigmoid(MLP(e))). The hierarchy is asserted as the source of 'supervised disentanglement', and Fig. 2 then validates disentanglement using the same intra-domain separation and cross-domain alignment that Eq. (3) already optimizes by cosine distance and L2 orthogonality. Moreover, since the sigmoid gates lie in (0,1) and f is a dot product, the mapping networks alone can satisfy Eq. (9) by shrinking e_hat_s toward zero and keeping e_hat_c near the anchor — encoder-level separation is not entailed, and gradients to e_s vanish as gates shrink.

full rationale

The paper's central claim — state-of-the-art cross-domain recommendation with up to 11.59% relative improvement (Table 3) — is an empirical result computed on a held-out 20% test split under standard Recall/HR/MRR/NDCG metrics against external baselines. There is no fit-then-predict pattern: BPR, encoder, item-contrastive, and decoder losses are training objectives, and none equals the evaluation metric. I checked all 41 references: none is authored by the present authors (Yuhan Wang, Qing Xie, Zhifeng Bao, Mengzi Tang, Lin Li, Yongjian Liu), so there is no self-citation chain and no uniqueness-imported-from-authors pattern. The hierarchical ordering in Eq. (9) is an acknowledged designer inductive bias, inspired by an external work (Han et al. [9]) and InfoNCE [25], not a derived theorem. The only mild self-referential element is the decoder 'supervision': the ranking is computed on the model's own transformed embeddings, so the 'supervisory signal' is an internal hierarchy, and Fig. 2's t-SNE largely re-demonstrates the separation/alignment that Eq. (3) already optimizes directly. This is an untested attribution (the sigmoid-gated mapping networks could absorb the ranking), but it does not make the reported accuracy equal to the input by construction. The manuscript honestly flags limitations: Sec. 4.7 admits 'the dependence on hyperparameter tuning may limit the practical applicability,' and footnote 3 asserts robustness under partial overlap without shown experiments; neither is circularity. Verdict: no significant circularity; the single minor self-referential component warrants score 2.

Assumptions & free parameters 10 free parameters · 5 assumptions · 0 invented entities

The method introduces no new entities; all innovation is in the loss design. However, the objective is assembled from at least 10 hand-tuned hyperparameters and several unproven inductive biases (the hierarchy ranking, orthogonality, and the suitability of the GNN embedding as an anchor). The reported gains depend on this tuning and these assumptions rather than on a self-contained derivation.

free parameters (10)
  • lambda_enc = 0.01 for Cloth-Elec and Book-Movie; 1.0 for Cloth-Sport
    Weight of the disentangled encoder loss, tuned from {0.01, 0.1, 1} per dataset pair (Sec. 4.7).
  • lambda_dec = 1.0 for Cloth-Elec; 0.01 for Cloth-Sport and Book-Movie
    Weight of the decoder contrastive loss, tuned jointly with lambda_enc (Sec. 4.7).
  • lambda_item = not reported in paper
    Weight of the item contrastive loss, selected from {0.01, 0.1, 1}.
  • temperature tau = within [0.05, 0.3]
    Temperature in InfoNCE-style losses, tuned with step 0.05.
  • L2 regularization coefficient = from {1e-3, 1e-4, 1e-5}
    Regularization strength, selected by fine-tuning.
  • dropout rate = from {0.1, 0.2, 0.3}
    Dropout probability, selected by fine-tuning.
  • learning rate = from {1e-3, 1e-4}
    Adam learning rate, selected by fine-tuning.
  • GCN depth H = 3
    Fixed to 3 with no sensitivity analysis, stated as capturing third-order connectivity.
  • batch size = 2048 for Sport&Cloth, 4096 for others
    Chosen by hardware capacity.
  • N-core threshold = N=10 for AmazonElec&Cloth and Douban; N=5 for AmazonSport&Cloth
    Data filtering thresholds chosen to ensure sufficient data after iterative sampling.
assumptions (5)
  • domain assumption Users share common interests across domains
    Motivates cross-domain transfer; stated throughout Sec. 1 and Sec. 3.2.
  • standard math GCN propagation (LightGCN-style) is a valid extractor of collaborative signals
    Equation (1) is adopted from prior work and treated as given, with no justification beyond common practice.
  • domain assumption InfoNCE contrastive loss provides a valid supervisory signal for representation ordering
    The decoder relies on pairwise ranking losses (Eq. 9); no formal guarantee or theoretical analysis is provided.
  • domain assumption Orthogonality between domain-shared and domain-specific features enforces semantic independence
    The encoder loss (Eq. 3) uses orthogonality; the paper's justification is empirical, not theoretical.
  • domain assumption Full user overlap suffices to evaluate the method
    The paper uses full-overlap datasets and asserts partial-overlap robustness without experiments (Sec. 4.1 footnote 3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Transferability and Consistency in Cross-Domain Recommendations via Supervised Disentanglement." pith.science (2026). https://pith.science/paper/VMNZFPDN

@misc{pith2026250717112,
  author       = {Pith},
  title        = {Pith review of: Enhancing Transferability and Consistency in Cross-Domain Recommendations via Supervised Disentanglement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VMNZFPDN}},
  note         = {Machine review of arXiv:2507.17112}
}
read the original abstract

Cross-domain recommendation (CDR) aims to alleviate the data sparsity by transferring knowledge across domains. Disentangled representation learning provides an effective solution to model complex user preferences by separating intra-domain features (domain-shared and domain-specific features), thereby enhancing robustness and interpretability. However, disentanglement-based CDR methods employing generative modeling or GNNs with contrastive objectives face two key challenges: (i) pre-separation strategies decouple features before extracting collaborative signals, disrupting intra-domain interactions and introducing noise; (ii) unsupervised disentanglement objectives lack explicit task-specific guidance, resulting in limited consistency and suboptimal alignment. To address these challenges, we propose DGCDR, a GNN-enhanced encoder-decoder framework. To handle challenge (i), DGCDR first applies GNN to extract high-order collaborative signals, providing enriched representations as a robust foundation for disentanglement. The encoder then dynamically disentangles features into domain-shared and -specific spaces, preserving collaborative information during the separation process. To handle challenge (ii), the decoder introduces an anchor-based supervision that leverages hierarchical feature relationships to enhance intra-domain consistency and cross-domain alignment. Extensive experiments on real-world datasets demonstrate that DGCDR achieves state-of-the-art performance, with improvements of up to 11.59% across key metrics. Qualitative analyses further validate its superior disentanglement quality and transferability. Our source code and datasets are available on GitHub for further comparison.

Figures

Figures reproduced from arXiv: 2507.17112 by the authors.

Figure 1
Figure 1. Overview of the proposed DGCDR. spaces. This symmetry design ensures structural consistency and supports effective feature interactions in the recommendations. To encode user preferences in each domain into two distinct rep￾resentations, we introduce constraints to ensure that the domain￾shared and domain-specific representations distinctly reflect com￾plementary aspects of user preferences. For domain-shared repre￾… view at source ↗
Figure 2
Figure 2. The t-SNE visualization of the disentangled intra [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 4
Figure 4. Performance comparison w.r.t. different values of [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 41 canonical work pages

  1. [1]

    Yu Bao, Hao Zhou, Shujian Huang, Lei Li, Lili Mou, Olga Vechtomova, Xin-Yu Dai, and Jiajun Chen. 2019. Generating Sentences from Disentangled Syntactic and Semantic Spaces. In Proceedings of the 57th Conference of the Association for Computational Linguistics. ACL, 6008–6019

    work page 2019

  2. [2]

    Jiangxia Cao, Xixun Lin, Xin Cong, Jing Ya, Tingwen Liu, and Bin Wang. 2022. DisenCDR: Learning Disentangled Representations for Cross-Domain Recom- mendation. In The 45th International ACM SIGIR Conference on Research and RecSys ’25, September 22–26, 2025, Prague, Czech Republic Wang et al. Development in Information Retrieval . ACM, 267–277

    work page 2022

  3. [3]

    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 38th IEEE International Conference on Data Engineering. IEEE, 2209–2223

    work page 2022

  4. [4]

    Pengyu Cheng, Martin Renqiang Min, Dinghan Shen, Christopher Malon, Yizhe Zhang, Yitong Li, and Lawrence Carin. 2020. Improving Disentangled Text Representation Learning with Information-Theoretic Guidance. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics . ACL, 7530–7541

    work page 2020

  5. [5]

    Emilien Dupont. 2018. Learning Disentangled Joint Continuous and Discrete Representations. In Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018 . 708–718

    work page 2018

  6. [6]

    Chen Gao, Yu Zheng, Nian Li, Yinfeng Li, Yingrong Qin, Jinghua Piao, Yuhan Quan, Jianxin Chang, Depeng Jin, Xiangnan He, and Yong Li. 2023. A Survey of Graph Neural Networks for Recommender Systems: Challenges, Methods, and Directions. Trans. Recomm. Syst. 1, 1 (2023), 1–51

    work page 2023

  7. [7]

    Huizhong Guo, Zhu Sun, Dongxia Wang, Tianjun Wei, Jinfeng Li, and Jie Zhang

    work page

  8. [8]

    Huizhong Guo, Dongxia Wang, Zhu Sun, Haonan Zhang, Jinfeng Li, and Jie Zhang. 2024. Configurable Fairness for New Item Recommendation Considering Entry Time of Items. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval . 437–447

    work page 2024

Show all 48 references
  1. [9]

    Teng-Yue Han, Pengfei Wang, Shaozhang Niu, and Chenliang Li. 2022. Modality Matches Modality: Pretraining Modality-Disentangled Item Representations for Recommendation. In The ACM Web Conference 2022. ACM, 2058–2066

    2022

  2. [10]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yong-Dong Zhang, and Meng Wang. 2020. LightGCN: Simplifying and Powering Graph Convolution Net- work for Recommendation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retriev...

    2020

  3. [11]

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

  4. [12]

    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. ACM, 1563–1572

    2019

  5. [13]

    Fengxin Li, Hongyan Liu, Jun He, and Xiaoyong Du. 2024. CausalCDR: Causal Embedding Learning for Cross-domain Recommendation. In Proceedings of the 2024 SIAM International Conference on Data Mining . SIAM, 553–561

    2024

  6. [14]

    Haoyang Li, Xin Wang, Ziwei Zhang, Zehuan Yuan, Hang Li, and Wenwu Zhu

  7. [15]

    Zihan Lin, Changxin Tian, Yupeng Hou, and Wayne Xin Zhao. 2022. Improving Graph Collaborative Filtering with Neighborhood-enriched Contrastive Learning. In The ACM Web Conference 2022. ACM, 2320–2329

    2022

  8. [16]

    Jing Liu, Lele Sun, Weizhi Nie, Peiguang Jing, and Yuting Su. 2024. Graph Disentangled Contrastive Learning with Personalized Transfer for Cross-Domain Recommendation. In Thirty-Eighth AAAI Conference on Artificial Intelligence . AAAI Press, 8769–8777

    2024

  9. [17]

    Meng Liu, Jianjun Li, Guohui Li, and Peng Pan. 2020. Cross Domain Recommen- dation via Bi-directional Transfer Graph Collaborative Filtering Networks. InThe 29th ACM International Conference on Information and Knowledge Management . ACM, 885–894

    2020

  10. [18]

    Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Rätsch, Sylvain Gelly, Bernhard Schölkopf, and Olivier Bachem. 2019. Challenging Common Assump- tions in the Unsupervised Learning of Disentangled Representations. In Proceed- ings of the 36th International Conference on M...

    2019

  11. [19]

    Tong Man, Huawei Shen, Xiaolong Jin, and Xueqi Cheng. 2017. Cross-Domain Recommendation: An Embedding and Mapping Approach. In Proceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence . ijcai.org, 2464–2470

    2017

  12. [20]

    Wentao Ning, Xiao Yan, Weiwen Liu, Reynold Cheng, Rui Zhang, and Bo Tang

  13. [21]

    Aghiles Salah, Thanh-Binh Tran, and Hady W. Lauw. 2021. Towards Source- Aligned Variational Models for Cross-Domain Recommendation. In Fifteenth ACM Conference on Recommender Systems . ACM, 176–186

    2021

  14. [22]

    Viégas, and Martin Wattenberg

    Daniel Smilkov, Nikhil Thorat, Charles Nicholson, Emily Reif, Fernanda B. Viégas, and Martin Wattenberg. 2016. Embedding Projector: Interactive Visualization and Interpretation of Embeddings. (2016). arXiv:1611.05469

    2016 arXiv

  15. [23]

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

    2024

  16. [24]

    Zhu Sun, Di Yu, Hui Fang, Jie Yang, Xinghua Qu, Jie Zhang, and Cong Geng

  17. [25]

    Aäron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation Learning with Contrastive Predictive Coding. abs/1807.03748 (2018). arXiv:1807.03748 http://arxiv.org/abs/1807.03748

    2018 arXiv

  18. [26]

    Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-SNE. Journal of machine learning research 9, 11 (2008), 2579–2605

    2008

  19. [27]

    Xin Wang, Hong Chen, Si’ao Tang, Zihao Wu, and Wenwu Zhu. 2024. Disentan- gled Representation Learning. IEEE Transactions on Pattern Analysis and Machine Intelligence 46, 12 (2024), 9677–9696

    2024

  20. [28]

    Xiang Wang, Hongye Jin, An Zhang, Xiangnan He, Tong Xu, and Tat-Seng Chua

  21. [29]

    Zongze Wu, Dani Lischinski, and Eli Shechtman. 2021. StyleSpace Analysis: Disentangled Controls for StyleGAN Image Generation. In IEEE Conference on Computer Vision and Pattern Recognition . Computer Vision Foundation / IEEE, 12863–12872

    2021

  22. [30]

    Kun Xu, Yuanzhen Xie, Liang Chen, and Zibin Zheng. 2021. Expanding Rela- tionship for Cross Domain Recommendation. In The 30th ACM International Conference on Information and Knowledge Management . ACM, 2251–2260

    2021

  23. [31]

    Zipeng Xu, Tianwei Lin, Hao Tang, Fu Li, Dongliang He, Nicu Sebe, Radu Timofte, Luc Van Gool, and Errui Ding. 2022. Predict, Prevent, and Evaluate: Disentangled Text-Driven Image Manipulation Empowered by Pre-Trained Vision-Language Model. In IEEE/CVF Conference on Computer Vi...

    2022

  24. [32]

    Mengyue Yang, Furui Liu, Zhitang Chen, Xinwei Shen, Jianye Hao, and Jun Wang

  25. [33]

    Tianzi Zang, Yanmin Zhu, Haobing Liu, Ruohan Zhang, and Jiadi Yu. 2023. A Survey on Cross-domain Recommendation: Taxonomies, Methods, and Future Directions. ACM Trans. Inf. Syst. 41, 2 (2023), 42:1–42:39

    2023

  26. [34]

    In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval

    Disentangled Graph Collaborative Filtering. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval. ACM, 1001–1010

  27. [35]

    Yu Zhang, Zhiyong Cheng, Fan Liu, Xun Yang, and Yuxin Peng. 2024. Decou- pled domain-specific and domain-conditional representation learning for cross- domain recommendation. Information Processing & Management 61, 2 (2024), 103689

    2024

  28. [36]

    Wayne Xin Zhao, Shanlei Mu, Yupeng Hou, Zihan Lin, Yushuo Chen, Xingyu Pan, Kaiyuan Li, Yujie Lu, Hui Wang, Changxin Tian, Yingqian Min, Zhichao Feng, Xinyan Fan, Xu Chen, Pengfei Wang, Wendi Ji, Yaliang Li, Xiaoling Wang, and Ji-Rong Wen. 2021. RecBole: Towards a Unified, Com...

    2021

  29. [37]

    Xin Zhou, Donghui Lin, Yong Liu, and Chunyan Miao. 2023. Layer-refined Graph Convolutional Networks for Recommendation. In 39th IEEE International Confer- ence on Data Engineering . IEEE, 1247–1259. https://doi.org/10.1109/ICDE55515. 2023.00100

    2023

  30. [38]

    Feng Zhu, Chaochao Chen, Yan Wang, Guanfeng Liu, and Xiaolin Zheng. 2019. DTCDR: A Framework for Dual-Target Cross-Domain Recommendation. In Pro- ceedings of the 28th ACM International Conference on Information and Knowledge Management. ACM, 1533–1542

    2019

  31. [39]

    In IEEE Conference on Computer Vision and Pattern Recognition

    CausalVAE: Disentangled Representation Learning via Neural Structural Causal Models. In IEEE Conference on Computer Vision and Pattern Recognition . Computer Vision Foundation / IEEE, 9593–9602

  32. [40]

    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 The 44th International ACM SIGIR Conference on Research and Development in Information Retri...

    2021

  33. [41]

    Ruohan Zhang, Tianzi Zang, Yanmin Zhu, Chunyang Wang, Ke Wang, and Jiadi Yu. 2023. Disentangled Contrastive Learning for Cross-Domain Recommendation. In Database Systems for Advanced Applications - 28th International Conference, Proceedings, Part II, Vol. 13944. Springer, 163–178

    2023

  34. [46]

    Jiajie Zhu, Yan Wang, Feng Zhu, and Zhu Sun. 2023. Domain Disentanglement with Interpolative Data Augmentation for Dual-Target Cross-Domain Recom- mendation. In Proceedings of the 17th ACM Conference on Recommender Systems (RecSys ’23). ACM, New York, NY, USA, 515–527

    2023

  35. [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 The Fifteenth ACM International Conference on Web Search and Data Mining . ACM, 1507–1515

    2022

  36. [2020]

    In Proceedings of the 14th ACM Conference on Recommender Systems

    Are We Evaluating Rigorously? Benchmarking Recommendation for Re- producible Evaluation and Fair Comparison. In Proceedings of the 14th ACM Conference on Recommender Systems . ACM, 23–32

  37. [2021]

    In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021

    Disentangled Contrastive Learning on Graphs. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021. 21872–21884

    2021

  38. [2023]

    In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management

    Multi-domain Recommendation with Embedding Disentangling and Do- main Alignment. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management . ACM, 1917–1927

    1917

  39. [2024]

    arXiv preprint arXiv:2403.03952 (2024)

    Bridging Language and Items for Retrieval and Recommendation. arXiv preprint arXiv:2403.03952 (2024)

    2024 arXiv

  40. [2025]

    In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Enhancing New-item Fairness in Dynamic Recommender Systems. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval . ACM, 1707–1716

Pith tools

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