Pith. sign in

REVIEW 4 major objections 5 minor 41 references

ATLAS: Learning to Recommend Across Unseen Domains

T0 review · 4 major / 5 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 ATLAS: a frozen recommender, trained on five source domains, transfers zero-shot to ten unseen domains and beats most strong baselines.

desk verdict RDG framing and architecture are promising, but the zero-shot claim is undercut by likely user overlap between source and target, and the GWLB theory is flawed. read the letter →

arxiv 2608.03899 v1 pith:J2R2SEB2 submitted 2026-08-04 cs.IR

Pervez Shaik , Prosenjit Biswas , Abhinav Thorat , Ravi Kolla , Niranjan Pedanekar This is my paper
classification cs.IR
keywords zero-shotrecommendationdomaingeneralizationGromov-Wassersteinalignmentadversarialadaptationresidualvectorquantizationcross-domainAmazonReviewsbenchmark
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

ATLAS is built on the hypothesis that heterogeneous recommendation domains share a common user–item interaction structure, even when users and catalogs are completely disjoint. The paper claims that a single model trained jointly on five Amazon domains—aligning item semantics, user-population geometry, and discrete codebooks—can be frozen and applied directly to ten domains it has never seen, with no target retraining and no language-model inference. On seven of ten unseen domains it reports higher HitRate than sequential, graph, cross-domain, quantization, and LLM-based baselines, with an average relative gain of 24% in HitRate. The result matters because it suggests transferable recommendation behavior can emerge from diverse interaction environments alone, rather than from target adaptation or large pretrained language models.

What carries the argument

The architecture combines three transfer mechanisms. Item-space unification: frozen Sentence-BERT text embeddings are projected by a shared MLP and made domain-indistinguishable by a gradient-reversal adversarial classifier. User-space unification: users are represented by concatenating a LightGCN collaborative embedding (used only in training) with the semantic mean of interacted items, then projected; the projected user populations are aligned across domains by minimizing the Gromov–Wasserstein lower bound (GWLB), which sorts all pairwise within-domain user cosine distances and compares the sorted lists across domain pairs. Residual vector quantization (RVQ) then compresses both user and i

What would settle it

Compute the squared 2-Wasserstein distance between the sorted pairwise user-distance distribution of each unseen target domain and the distribution the source domains were aligned to. If zero-shot HitRate does not fall as this distance grows—or better, if ATLAS's advantage over BPR-only does not vanish when the target distance distribution is deliberately perturbed—then the geometric-alignment claim is false.

Watch

Extended reading notes

Core claim

The paper's central claim is that recommendation knowledge is not domain-bound: if source-specific correlations are suppressed while the common structure of how users relate to items is preserved, a scoring function learned from multiple heterogeneous source domains remains usable in an unseen domain. ATLAS operationalizes this by learning a shared, domain-invariant user-item retrieval space from five disjoint Amazon source domains, then freezing the projections and codebooks. At test time, target users are represented only through the semantic mean of their interaction history, target items through their text descriptions, and retrieval is a dot product between codebook-reconstructed vector

Load-bearing premise

The transfer works only if users in an unseen catalog have roughly the same distribution of pairwise similarities as the training domains; if a new domain's user population has different relational geometry, the frozen alignment transfers little.

Editorial extensions

If this is right

  • A recommender can be trained once on a diverse set of catalogs and deployed to a new catalog with only item text and user history, no target labels, no retraining, and no LLM call at serve time.
  • The 24% average relative HitRate gain, if it holds across broader benchmarks, would make zero-shot retrieval competitive with adaptation-based methods for cold-start catalog launches.
  • Source-domain diversity becomes a concrete data-acquisition guideline: because zero-shot transfer improves monotonically as more heterogeneous source domains are added, diversity should be prioritized over source-domain scale.
  • The codebook reconstruction outperforming continuous projections indicates that discretization itself is a transfer mechanism, not just compression, because it snaps target embeddings onto source-supported regions of the sphere.
  • Domain invariance does not come at the cost of in-domain accuracy: ATLAS remains competitive on the five source domains, so the alignment losses do not destroy ranking signal.

Reading between the lines

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

  • A natural testable extension is to replace the frozen Sentence-BERT encoder with a multilingual or domain-agnostic text encoder; the alignment and codebook machinery should then transfer to non-English catalogs without architectural change.
  • Because only the semantic channel is available for target users, a stronger identity-free aggregation over interaction history (for example, attention over item embeddings rather than a mean pool) could improve zero-shot precision without violating the frozen-model constraint.
  • GWLB matches only the marginal distribution of pairwise distances, not the full transport plan; two target user populations with identical distance histograms but different cluster topologies could evade the alignment, so a per-domain diagnostic based on the distance between source and target distance distributions would be a practical guard.
  • The 24% average gain is not uniform: on some unseen domains ATLAS underperforms baselines, so a deployment would need cheap domain-similarity triggers to decide when to trust zero-shot retrieval versus fall back to a domain-specific model.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper introduces ATLAS, a multi-source recommendation framework for the proposed Recommendation Domain Generalization (RDG) setting, in which a frozen model is trained on several source domains and deployed to target domains whose users and items are disjoint from those seen in training. ATLAS combines a shared item projection trained with an adversarial domain-discrimination loss, a user-space alignment based on the Gromov–Wasserstein lower bound (GWLB) between pairwise distance distributions, and residual vector quantization (RVQ) codebooks. Experiments on five Amazon source domains and ten unseen target domains report that ATLAS outperforms sequential, graph-based, cross-domain, quantization-based, and LLM-based baselines on a majority of target domains, with an average relative HitRate improvement of 24%. Ablations and representation analyses are also presented.

Significance. If the empirical claims hold, the paper would make a strong contribution: it demonstrates that a single frozen retrieval model, trained without target data or LLM inference, can transfer to previously unseen recommendation domains. The evaluation is unusually broad for this setting, with five source and ten target domains, full-ranking leave-one-out evaluation, ablations over the main loss components, and multi-seed stability checks. The codebook-utilization diagnostics and the explicit RDG formulation are also useful. However, two load-bearing issues currently prevent full confidence: the theoretical justification for the user-alignment loss contains an invalid implication, and the central 'disjoint users' assumption is not verified or enforced in the experimental pipeline. Both are fixable in principle, but they require substantive changes to the manuscript and possibly to the experimental results.

major comments (4)
  1. [Appendix, Corollary 1 and Eq. (14)] The claimed chain Lgw→0 ⇒ GWLB2→0 ⇒ SLB2→0 ⇒ D2→0 is logically invalid. Proposition 1 states SLB2 ≤ D2, and Eq. (13) states SLB2^2 ≥ GWLB2. From GWLB2→0 one can conclude only SLB2→0, which is a lower bound on D2; D2 can remain positive while the lower bound vanishes. Therefore the conclusion that the source mm-spaces become isomorphic (Theorem 1) does not follow, and the statement that Lgw is 'provably necessary' toward isomorphism is unsupported. Since this corollary is the formal justification for the user-unification loss, the theoretical grounding of Lgw needs to be corrected, e.g., by reframing GWLB as a heuristic or weaker necessary condition rather than a certificate of geometric equivalence.
  2. [Experimental Setup / Table 9] The RDG problem formulation states that source and target users and items are disjoint, but the experimental setup only removes users with fewer than five interactions. No step is described that removes reviewer IDs appearing in both source and target domains or that deduplicates users across the five source domains. Table 9 reports per-domain user counts whose totals are exactly the sum of the column values (e.g., source total 2,803,448 = 588,804 + 489,076 + 726,286 + 117,742 + 881,540), which is what one would observe if shared Amazon reviewer IDs were counted separately per domain. If a target test user also has interactions in a source domain, ATLAS has been trained on that user's source-domain interactions through the user projection, BPR loss, and GW alignment, violating the 'disjoint users' premise. This directly threatens the zero-shot claim. Please enforce and verify user- and i
  3. [Abstract / Table 3] The headline 'average relative gain in HitRate of 24%' is not reproducible from the reported numbers as stated. For example, relative to UniSRecZS, the per-domain HR@10 gains in Table 3 average far above 24% (about 53%), while relative to other baselines the figure would differ. The paper must specify the reference baseline and the exact aggregation formula used for the average relative gain. Additionally, the table layout mixes zero-shot and fine-tuned rows; the text says GWCDR, PrepRec, and LLM-RecG are trained separately for each source–target pair, so they are not zero-shot. Clarify which rows belong to the zero-shot comparison set and ensure the 'wins on seven of ten target domains' claim refers only to genuinely zero-shot baselines.
  4. [Zero-Shot Inference (Projection with frozen encoders)] At inference the target user is represented by [0; e_sem], while during training the user projection receives [e_c; e_sem]. The main ATLASZS results were obtained without the modality-dropout variant, even though Appendix Table 7 shows that the dropout variant improves zero-shot performance on several domains. The paper should state explicitly which variant is used in Table 3 and Appendix Table 13, and discuss why the model that is trained and evaluated under the same input distribution is not the default configuration. As it stands, the reported zero-shot numbers are from a model experiencing an input-distribution shift that the authors themselves identify and mitigate only in an ablation.
minor comments (5)
  1. [Table 3] There is a typo in the ATLASFT row: '0.0144.0070' should presumably be '0.0144 0.0070'. Also, 'Musi.' is used inconsistently for Musical Instruments; consider a consistent column abbreviation.
  2. [RQ1 / Wilcoxon tests] The footnote reports p=0.13 and p=0.05 for ATLASZS vs UniSRecZS without stating the hypothesis or whether the test is one- or two-sided. Please clarify the exact comparison and the direction of the alternative.
  3. [Figure 1] Figure 1 contains typos: 'Gromov-Wasserstien' and 'Gromov-W a ss er s tien' should be 'Gromov-Wasserstein'. These typos also appear in the caption text.
  4. [Experimental Setup / Baselines] The adaptation of same-domain baselines by 'jointly training on a unified interaction pool' is described only briefly. Please provide details of the merged training protocol, since this affects the credibility of the in-domain comparisons in Table 2.
  5. [Table 9] Even if user disjointness turns out to be satisfied, the table should state how overlapping IDs were handled. Merely summing per-domain counts suggests no deduplication was performed, and a short sentence in the setup would remove ambiguity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central zero-shot claim is evaluated against external baselines and the losses are defined independently of the target-domain evaluation.

full rationale

ATLAS's central claim—that a frozen model trained on five source domains can zero-shot recommend on ten unseen domains—is tested against external baselines (Table 3) under a fixed protocol, so the reported gains are not constructed from the training objective. The losses L_bpr, L_adv, L_gw, and L_rvq are defined independently of the evaluation metrics (HR@10/NDCG@10) and of the target-domain data. No parameter is fitted to target interactions and then reported as a prediction; the only target information used at inference is the user's history to construct a query representation, which is not optimized. The GWLB alignment is explicitly borrowed from Mémoli and Peyré-Cuturi, not from the authors' own prior work, so there is no load-bearing self-citation or imported uniqueness theorem. The appendix's Corollary 1 chain (GWLB→0 ⇒ SLB→0 ⇒ D2→0) is invalid because the cited inequalities give SLB2 ≤ D2 and SLB2^2 ≥ GWLB2, making GWLB only a lower bound on D2; however, this is a proof-level correctness flaw, not a circular definition or a fitted-input-renamed-as-prediction. Similarly, the skeptic's user-overlap concern is a potential data-leakage threat to external validity, not a circularity of the derivation. No equation in the paper reduces a prediction to a fitted input by construction, and no central premise is justified solely by a self-citation. Therefore the appropriate circularity finding is no significant circularity.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central claim rests on assumption of transferable interaction structure, a frozen SBERT semantic space, and the transfer of user distance geometry to unseen domains. The method adds no new physical or conceptual entity; all building blocks are prior techniques. Hyperparameters and loss weights are manually chosen and are part of the load-bearing configuration.

free parameters (8)
  • λ_bpr = 1.7
    Weight of the sampled-softmax recommendation loss; set by hand, affects ranking-signal balance.
  • λ_adv = 2.0
    Weight for item adversarial loss; chosen by hand; item-domain invariance strength.
  • λ_gw = 2.0
    Weight for Gromov-Wasserstein lower bound user alignment; central to user-space transfer claim.
  • λ_commit, λ_book, λ_ent = 0.25, 2.0, 0.05
    RVQ commitment, codebook, and entropy penalty weights; prevent codebook collapse.
  • λ_centroid, λ_user_div, λ_var = 2.0, 1.0, 1.0
    Auxiliary stabilizers; claimed to not be contributions but affect training.
  • RVQ codebook sizes / levels = L=3; sizes 512,1024,2048 per level
    Capacity of discrete latent space; chosen by hand, affects information bottleneck.
  • softmax temp, VQ temp, Sinkhorn iter/reg = 0.20, 0.25, 5, 0.05
    Smoothing/quantization annealing hyperparameters.
  • number of negatives J = not reported
    Sampled-softmax denominator size; missing from hyperparameters, could affect ranking quality.
assumptions (6)
  • domain assumption Heterogeneous recommendation domains share transferable user-item interaction structure (the central hypothesis of RDG).
    Stated in Introduction and Conclusion; not derived; the paper depends on it.
  • domain assumption Frozen Sentence-BERT provides a common semantic item space across domains.
    Item-space unification assumes text embeddings align semantically across catalogues.
  • domain assumption Unseen target users have the same internal pairwise-distance geometry as aligned source domains.
    Stated in Appendix after Equation 8: 'an unseen target domain with similar preference geometry then naturally inherits it'; never measured on targets.
  • standard math Mémoli's SLB2/GW distance theorems and the quantile-integral identity hold as cited.
    Borrowed from Mémoli 2011 and Peyré & Cuturi 2019; the paper applies them with incorrect implication direction.
  • domain assumption Ben-David divergence bounds apply to projected item embeddings for ranking tasks.
    Appendix Lemma 1/Theorem 2; the theory is for classification, adapted to retrieval without proof of the adaptation.
  • domain assumption LightGCN pretrained collaborative embeddings are useful when projected through a shared user projection.
    User representations concatenate LightGCN and semantic embeddings; at inference the LightGCN channel is zero-imputed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ATLAS: Learning to Recommend Across Unseen Domains." pith.science (2026). https://pith.science/paper/J2R2SEB2

@misc{pith2026260803899,
  author       = {Pith},
  title        = {Pith review of: ATLAS: Learning to Recommend Across Unseen Domains},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J2R2SEB2}},
  note         = {Machine review of arXiv:2608.03899}
}
read the original abstract

Recommender systems remain domain-bound: a model trained on one interaction environment typically requires retraining or target-domain adaptation before it can operate on a new catalogue. A recommender trained on movies cannot be directly deployed to recommend groceries or video games. Existing approaches mitigate this by transferring restricted forms of recommendation knowledge, adapting to the target domain, or leveraging large language models (LLMs) for transferable representations. We instead ask whether recommendation-specific knowledge learned solely from multiple heterogeneous domains can generalize to entirely unseen domains without target-domain adaptation or language-model pretraining. We introduce ATLAS, a multi-source recommendation domain generalization framework that learns a shared, domain-invariant user-item representation from disjoint source domains, enabling zero-shot recommendation on unseen domains. ATLAS combines a Gromov-Wasserstein alignment that preserves how users relate to one another across domains, an adversarial objective that makes item representations indistinguishable across domains, and residual vector quantization (RVQ) codebooks that compress user and item embeddings into a discrete latent space, capturing hierarchical interaction patterns while suppressing domain-specific variation. Trained on five Amazon domains and applied directly to ten unseen domains, ATLAS outperforms state-of-the-art sequential, graph-based, cross-domain, quantization-based, and LLM-based baselines on most unseen domains, with an average relative gain in HitRate of 24%. Ablations and representation analyses validate each component, and we identify a pronounced source-domain diversity effect: increasing source heterogeneity substantially improves zero-shot transfer. ATLAS establishes recommendation domain generalization as a promising paradigm for zero-shot recommendation.

Figures

Figures reproduced from arXiv: 2608.03899 by the authors.

Figure 1
Figure 1. ATLAS architecture N = n 2  strictly upper-triangular entries into the empirical distance measure µ k = 1 N P a<b δDk ab . Since µ a and µ b are one-dimensional empirical measures of equal size, their squared 2-Wasserstein distance reduces to a sum over order statistics(Peyré and Cuturi 2019): GWLB2 (a, b) = W2 2 (µ a , µ b ) = 1 N XN r=1 d a (r) − d b (r) 2 , where d k (r) is the r-th smallest pairwise distance … view at source ↗
Figure 2
Figure 2. User embedding alignment Auto. Elec. Beauty Movies Games Auto. Elec. Beauty Movies Games 0.00 1.85 1.95 1.99 1.96 1.85 0.00 1.96 1.99 1.88 1.95 1.96 0.00 1.97 1.98 1.99 1.99 1.97 0.00 1.84 1.96 1.88 1.98 1.84 0.00 Epoch 1 Auto. Elec. Beauty Movies Games Auto. Elec. Beauty Movies Games 0.00 0.64 0.69 1.08 0.60 0.64 0.00 0.73 0.92 0.53 0.69 0.73 0.00 0.99 0.62 1.08 0.92 0.99 0.00 0.98 0.60 0.53 0.62 0.98 0.00 Epoch 40… view at source ↗
Figure 5
Figure 5. Codebook assignment frequency histograms for user (top) and item (bottom) codebooks, across levels [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figures from the paper (1 more)
Figure 6
Figure 6. Figure 6: Proxy A-distance between source-domain item representations at epochs 1, 100, 500, 1000, 2000, and 4000 [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 19 canonical work pages

  1. [1]

    Proceedings of the 32nd International Conference on Machine Learning , pages =

    Unsupervised Domain Adaptation by Backpropagation , author =. Proceedings of the 32nd International Conference on Machine Learning , pages =. 2015 , editor =

    work page 2015

  2. [2]

    Machine learning , volume=

    A theory of learning from different domains , author=. Machine learning , volume=. 2010 , publisher=

    2010

  3. [3]

    2014 , eprint=

    Generative Adversarial Networks , author=. 2014 , eprint=

    2014

  4. [4]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Multi-adversarial domain adaptation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    work page

  5. [5]

    Gromov–Wasserstein Distances and the Metric Approach to Object Matching , volume =

    Mémoli, Facundo , year =. Gromov–Wasserstein Distances and the Metric Approach to Object Matching , volume =. Foundations of Computational Mathematics , publisher =. doi:10.1007/s10208-011-9093-5 , number =

    doi:10.1007/s10208-011-9093-5

  6. [6]

    Findings of the Association for Computational Linguistics: NAACL 2024 , pages=

    Llm-rec: Personalized recommendation via prompting large language models , author=. Findings of the Association for Computational Linguistics: NAACL 2024 , pages=

    work page 2024

  7. [7]

    Studies in Applied Mathematics , volume =

    Mémoli, Facundo and Needham, Tom , title =. Studies in Applied Mathematics , volume =. doi:https://doi.org/10.1111/sapm.12526 , url =. https://onlinelibrary.wiley.com/doi/pdf/10.1111/sapm.12526 , year =

    doi:10.1111/sapm.12526

  8. [8]

    Advances in neural information processing systems , volume=

    Neural discrete representation learning , author=. Advances in neural information processing systems , volume=

Show all 41 references
  1. [9]

    Gromov-wasserstein averaging of kernel and distance matrices , year =

    Peyr\'. Gromov-wasserstein averaging of kernel and distance matrices , year =. Proceedings of the 33rd International Conference on International Conference on Machine Learning - Volume 48 , pages =

  2. [10]

    2019 , publisher=

    Computational optimal transport: With applications to data science , author=. 2019 , publisher=

    2019

  3. [11]

    Sinkhorn Distances: Lightspeed Computation of Optimal Transport , url =

    Cuturi, Marco , booktitle =. Sinkhorn Distances: Lightspeed Computation of Optimal Transport , url =

  4. [12]

    Yupeng Hou and Zhankui He and Julian McAuley and Wayne Xin Zhao , title =

  5. [13]

    2022 , isbn =

    Li, Xinhang and Qiu, Zhaopeng and Zhao, Xiangyu and Wang, Zihao and Zhang, Yong and Xing, Chunxiao and Wu, Xian , title =. 2022 , isbn =. doi:10.1145/3511808.3557338 , booktitle =

    2022

  6. [14]

    2018 IEEE international conference on data mining (ICDM) , pages=

    Self-attentive sequential recommendation , author=. 2018 IEEE international conference on data mining (ICDM) , pages=. 2018 , organization=

    2018

  7. [15]

    arXiv preprint arXiv:2403.03952 , year=

    Bridging Language and Items for Retrieval and Recommendation , author=. arXiv preprint arXiv:2403.03952 , year=

    arXiv

  8. [16]

    2023 IEEE 39th International Conference on Data Engineering (ICDE) , pages=

    Robust attributed graph alignment via joint structure learning and optimal transport , author=. 2023 IEEE 39th International Conference on Data Engineering (ICDE) , pages=. 2023 , organization=

    2023

  9. [17]

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

    Lightgcn: Simplifying and powering graph convolution network for recommendation , author=. Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval , pages=

  10. [18]

    Proceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval , pages=

    Neural graph collaborative filtering , author=. Proceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval , pages=

  11. [19]

    Proceedings of the 28th ACM international conference on information and knowledge management , pages=

    BERT4Rec: Sequential recommendation with bidirectional encoder representations from transformer , author=. Proceedings of the 28th ACM international conference on information and knowledge management , pages=

  12. [20]

    Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining , pages=

    Towards universal sequence representation learning for recommender systems , author=. Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining , pages=

  13. [21]

    Proceedings of the 18th ACM Conference on Recommender Systems , pages=

    A pre-trained zero-shot sequential recommendation framework via popularity dynamics , author=. Proceedings of the 18th ACM Conference on Recommender Systems , pages=

  14. [22]

    Proceedings of the 17th ACM conference on recommender systems , pages=

    Tallrec: An effective and efficient tuning framework to align large language model with recommendation , author=. Proceedings of the 17th ACM conference on recommender systems , pages=

  15. [23]

    Proceedings of the Nineteenth ACM Conference on Recommender Systems , pages=

    Llm-recg: A semantic bias-aware framework for zero-shot sequential recommendation , author=. Proceedings of the Nineteenth ACM Conference on Recommender Systems , pages=

  16. [24]

    2025 , issue_date =

    Xin, Haoran and Sun, Ying and Wang, Chao and Xiong, Hui , title =. 2025 , issue_date =. doi:10.1145/3715099 , journal =

    2025 doi

  17. [25]

    arXiv preprint arXiv:2105.08318 , year=

    Zero-shot recommender systems , author=. arXiv preprint arXiv:2105.08318 , year=

    arXiv

  18. [26]

    Advances in Neural Information Processing Systems , volume=

    Recommender systems with generative retrieval , author=. Advances in Neural Information Processing Systems , volume=

  19. [27]

    arXiv preprint arXiv:2510.12369 , year=

    A Hierarchical Quantized Tokenization Framework for Task-Adaptive Graph Representation Learning , author=. arXiv preprint arXiv:2510.12369 , year=

    arXiv

  20. [28]

    Sentence-bert: Sentence embeddings using siamese bert-networks , author=. Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP) , pages=

    2019

  21. [29]

    arXiv preprint arXiv:1511.06939 , year=

    Session-based recommendations with recurrent neural networks , author=. arXiv preprint arXiv:1511.06939 , year=

    arXiv

  22. [30]

    Proceedings of the 27th ACM international conference on information and knowledge management , pages=

    Conet: Collaborative cross networks for cross-domain recommendation , author=. Proceedings of the 27th ACM international conference on information and knowledge management , pages=

  23. [31]

    , author=

    Cross-domain recommendation: An embedding and mapping approach. , author=. Ijcai , volume=

  24. [32]

    Proceedings of the 24th International Conference on World Wide Web (WWW) , pages =

    Elkahky, Ali Mamdouh and Song, Yang and He, Xiaodong , title =. Proceedings of the 24th International Conference on World Wide Web (WWW) , pages =. 2015 , doi =

    2015

  25. [33]

    Proceedings of the 31st ACM International Conference on Information & Knowledge Management , pages=

    Gromov-wasserstein guided representation learning for cross-domain recommendation , author=. Proceedings of the 31st ACM International Conference on Information & Knowledge Management , pages=

  26. [34]

    Proceedings of the 45th International ACM SIGIR conference on research and development in information retrieval , pages=

    Exploiting variational domain-invariant user embedding for partially overlapped cross domain recommendation , author=. Proceedings of the 45th International ACM SIGIR conference on research and development in information retrieval , pages=

  27. [35]

    Pacific-Asia Conference on Knowledge Discovery and Data Mining , pages=

    Modeling User Preferences as Distributions for Optimal Transport-Based Cross-Domain Recommendation under Non-overlapping Settings , author=. Pacific-Asia Conference on Knowledge Discovery and Data Mining , pages=. 2026 , organization=

    2026

  28. [36]

    Proceedings of the 17th ACM International Conference on Web Search and Data Mining , pages=

    Pre-trained recommender systems: A causal debiasing perspective , author=. Proceedings of the 17th ACM International Conference on Web Search and Data Mining , pages=

  29. [37]

    Proceedings of the 16th ACM conference on recommender systems , pages=

    Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5) , author=. Proceedings of the 16th ACM conference on recommender systems , pages=

  30. [38]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Recbase: Generative foundation model pretraining for zero-shot recommendation , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  31. [39]

    Proceedings of the 16th ACM Conference on Recommender Systems , pages =

    Geng, Shijie and Liu, Shuchang and Fu, Zuohui and Ge, Yingqiang and Zhang, Yongfeng , title =. Proceedings of the 16th ACM Conference on Recommender Systems , pages =. 2022 , isbn =. doi:10.1145/3523227.3546767 , abstract =

    2022

  32. [40]

    ACM Trans

    Bao, Keqin and Zhang, Jizhi and Wang, Wenjie and Zhang, Yang and Yang, Zhengyi and Luo, Yanchen and Chen, Chong and Feng, Fuli and Tian, Qi , title =. ACM Trans. Recomm. Syst. , month = apr, articleno =. 2025 , issue_date =. doi:10.1145/3716393 , abstract =

    2025 doi

  33. [41]

    Proceedings of the 32nd International Conference on Machine Learning , pages =

    Learning Transferable Features with Deep Adaptation Networks , author =. Proceedings of the 32nd International Conference on Machine Learning , pages =. 2015 , editor =

    2015

Pith tools

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