Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Measure Domain's Gap: A Similar Domain Selection Principle for Multi-Domain Recommendation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper proposes a Similar Domain Selection Principle (SDSP) that, layer on top of existing multi-domain recommender systems, masks out experts from dissimilar domains to prevent negative transfer and improve click-through-rate…

desk verdict SDSP is a legitimate incremental contribution to multi-domain recommendation, but the unproven prefix-only search assumption is the real soft spot and the reporting needs tightening before it can be trusted. read the letter →

arxiv 2505.20227 v1 pith:WK7UR7T3 submitted 2025-05-26 cs.IR

Yi Wen , Yue Liu , Derong Xu , Huishi Luo , Pengyue Jia , Yiqing Wu , Siwei Wang , Ke Liang
show 4 more authors
Maolin Wang Yiqi Wang Fuzhen Zhuang Xiangyu Zhao
This is my paper · ORCID
classification cs.IR
keywords multi-domainrecommendationnegativetransferdomainselectionprototypelearningepsilon-greedybanditclick-throughratepredictionmixture-of-experts
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

Multi-domain recommender systems transfer knowledge across domains, but when domains conflict or one is low quality, sharing everything hurts performance — the negative transfer problem. This paper claims SDSP, a lightweight plug-in selection principle, fixes that by measuring the gap between domains and, for each target domain, choosing a similar-domain set whose experts are allowed to contribute. Applying SDSP to MMOE and PLE lifts overall AUC on Movielens from 0.8077 to 0.8151 and from 0.8085 to 0.8147, with similar gains on Amazon and Douban. The principle is the first to explicitly measure domain-level gaps and select domains dynamically in multi-domain recommendation, and it adds only about 0.4 ms of inference time per batch.

What carries the argument

The prototype-based domain distance measure: each domain's learned embeddings are compressed into $M$ prototypes by an encoder, the distance from domain $d_1$ to $d_2$ is the average distance from each prototype of $d_1$ to its nearest prototype in $d_2$, computed with the L2 norm. It is asymmetric, cheap ($O(BM+M^2)$ rather than $O(B^2)$), and model-aware because the prototypes come from the model's own representations. On top of this, the selection is a decaying epsilon-greedy bandit: at regular iterations the algorithm either randomly explores a prefix of the distance-sorted candidate list or exploits the domain metric (AUC) recorded for each state, then masks out experts whose domains are not in the chosen set.

What would settle it

Run MMOE+SDSP on a three-domain setup where the middle-distance domain is harmful but the farthest domain is helpful (for example, a synthetic mixture with known transfer conflicts), and compare the prefix-only search against exhaustive search over all subsets; if exhaustive search finds a non-prefix set with higher validation AUC, the prefix assumption is refuted.

Watch

Extended reading notes

Core claim

The central claim is that negative transfer in multi-domain recommendation can be alleviated not by redesigning the shared structure but by choosing which domains' experts each target domain may draw on. SDSP learns a prototype per domain from the model's own representations, defines an asymmetric prototype-based distance between domains, and uses an epsilon-greedy bandit over the sorted distance prefixes to pick the beneficial domain set; experts outside that set are masked out with their gate weights set to $-\infty$. The paper shows this selection loop consistently beats fixed-sharing baselines (MMOE, PLE, STAR, SAR-Net, Adasparse, AdaptDHM, M2M, PPNet, EPNet) on Movielens, Amazon, and Douban in AUC and LogLoss, with statistically significant gains, and that plugging SDSP into MMOE, PLE, and SAR-Net improves almost every per-domain and overall AUC. A theorem bounds the loss gap: if a target's embedding depends on the selected domains, the best attainable cross-entropy loss exceeds the full-data optimum by at least $\Delta^i_p = I(X_i; Y_i) - \min_{j\in S_i} I(X_j; Y_i)$.

Load-bearing premise

The load-bearing premise is that the best set of domains to share with a target is always a prefix of the list sorted by measured distance: if a farther domain is helpful but a nearer one is harmful, the prefix-only search can miss the best combination.

Editorial extensions

If this is right

  • Plugging SDSP into MMOE, PLE, and SAR-Net yields statistically significant AUC and LogLoss gains across Movielens, Amazon, and Douban; for example, Movielens overall AUC rises from 0.8077 to 0.8151 for MMOE+SDSP.
  • The method is lightweight: inference time increases by only 0.38–0.55 ms per batch of 4096, and parameter count grows by about 86–90K for MMOE and PLE.
  • SDSP reduces negative transfer: on Movielens, PLE+SDSP avoids the SENIOR→YOUNG and YOUNG→SENIOR deteriorations that plain PLE exhibits.
  • Both components matter: the ablation shows that the prototype-based distance plus the dynamic selection outperforms random selection and an average-representation distance on Movielens.
  • The distance-sorted prefix restriction cuts the candidate set from $O(2^{D-1})$ to $O(D)$, making the combinatorial search tractable.

Reading between the lines

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

  • Because the prototype distance is asymmetric, SDSP should also suit settings with highly asymmetric user overlap, such as cross-platform behavior transfer; a natural test is to apply it to heterogeneous domain pairs and check whether the chosen direction of sharing flips when the domains are swapped.
  • The nested-prefix assumption is the main risk: if a farther domain is beneficial while a nearer one is harmful, prefix-only search can miss the optimal set; a testable extension is to relax the search to a beam of top-$k$ candidate sets and measure the AUC gain.
  • Since the prototypes come from the model's own representations, the selected domain sets should shift as the model's learning capacity changes during training; logging the chosen sets over epochs would verify whether the selection is genuinely dynamic.
  • The same selection principle could be ported to other multi-task sharing architectures where negative transfer appears, such as multi-task learning with shared backbones, by treating each task as a domain and reusing the prototype-distance bandit.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes SDSP (Similar Domain Selection Principle) for multi-domain recommendation (MDR). The method learns a trainable prototype per domain, defines an asymmetric prototype-based distance between domains, and then, for each target domain, uses an epsilon-greedy bandit over a restricted set of candidate domain subsets to decide which other domains' experts to mask. The mask is applied to the gates of an MMOE-style shared-expert architecture, and the overall loss adds a prototype reconstruction term. The authors report that incorporating SDSP into MMOE, PLE, and SAR-Net improves AUC/LogLoss on Movielens, Amazon, and Douban, with small inference-time and parameter overhead. A theoretical section attempts to bound the information gap caused by negative transfer, and experiments include hyperparameter analysis, ablations, negative-transfer visualization, and efficiency measurements.

Significance. If the claims hold, SDSP would be a useful lightweight plug-in for MDR: it directly targets negative transfer, is compatible with several existing backbones, and its prototype-based distance is computationally attractive. The paper's main positive features are the clean plug-in formulation, the asymmetric distance definition, the inclusion of an efficiency analysis, and the attempt to ablate the distance measure and selection strategy. The reported gains are consistent in direction across datasets and backbones. However, the central conceptual claim -- that SDSP selects beneficial similar domains based on a learned distance -- is currently not fully supported: the candidate set is artificially restricted to prefixes of the distance ordering (Remark 2), the selection reward is the validation AUC, and the empirical reporting lacks error bars or seed-level statistics despite t-test claims. These issues are load-bearing for the paper's interpretation, but they are addressable within the scope of a revision.

major comments (4)
  1. [Section 3.3, Remark 2] The restriction to prefix candidate sets is unproven and appears to contradict the paper's own motivation. With domains ordered by distance as [target, A, B], the set {target, B} is never considered, yet Figure 1 shows asymmetric transfer (SENIOR hurts YOUNG while YOUNG helps SENIOR for M2M), so a closer domain can be harmful and a farther domain can be beneficial. The claim that the search space reduces from O(2^(D-1)) to O(D) is therefore not a harmless efficiency shortcut unless a nested-prefix property is proved. Since the bandit can only choose among prefixes, it cannot repair a bad distance-based ordering. I request either a proof of the prefix property or an experiment on the D=3 datasets comparing SDSP against exhaustive selection over all subsets; without this, the central claim that SDSP selects suitable domains is not established.
  2. [Section 5.2, Table 2] The numerical claims in the text do not match Table 2. The paper states that MMOE+SDSP exceeds SharedBottom by 1.26%, 3.86%, and 0.32% in overall AUC, but the corresponding differences in Table 2 are 0.0127, 0.0049, and 0.0032 (i.e., about 1.58%, 0.69%, and 0.40% relative), respectively. The 3.86% figure is especially far off. Additionally, the tables report only point estimates, with no standard deviations, confidence intervals, or number of seeds, yet the text claims statistically significant improvements via two-sided t-tests with p<0.05. These significance claims are unsupported as reported. The authors should correct the arithmetic and provide seed-level results or justify why a t-test is valid without them.
  3. [Section 3.4, Algorithm 1] Algorithm 1 is underspecified in ways that affect reproducibility. The action space is not defined: "Randomly select a similar domain set for S_d" could mean sampling uniformly from all subsets, from prefixes of the distance ordering, or from some other family. The value-function update is described only as "Update value function of state S_d using metric r_d"; the paper does not state the update rule, how the reward is aggregated over epochs, or how ties are broken. The selection interval l is introduced but its default value is not given in the experimental setup. Please specify the exact state space, action space, reward assignment, update equation, and default values.
  4. [Section 3.4 and Section 5.5, Figure 4] The evaluation metric used as the bandit reward is the validation AUC, and the final reported metric is the test AUC. Part of the observed improvement may therefore reflect validation-based model selection rather than the predictive quality of the prototype distance. The ablation in Figure 4 shows that a poor distance measure ("Average") still helps relative to random selection, which is consistent with this concern. To isolate the contribution of the unsupervised distance, the authors should compare SDSP against a variant that runs the same epsilon-greedy selection over validation AUC without any distance-based ordering (or over all subsets), and report the selected domain sets. This would clarify whether the chosen distance actually identifies beneficial domains or whether the bandit alone is responsible for the gains.
minor comments (5)
  1. [Section 5, RQ1] The text contains the typo "exsiting" in RQ1; it should be "existing."
  2. [Figure 3] The hyperparameter figure contains many unreadable character tokens (e.g., Unicode escape sequences) in its axis labels and legends; the figure needs to be regenerated with proper text.
  3. [Section 5.6, Figure 5] The sentence "when the domain SENIOR causes a strong negative transfer to the domain SENIOR" appears to contain an error: the text likely means negative transfer to YOUNG or from YOUNG to SENIOR. Please rephrase for clarity.
  4. [Section 4, Theorem 3] The assumption Z_i = g_j(X_j) for all j in S_i is unclear when S_i contains more than one domain, since it would require all g_j(X_j) to be equal to the same Z_i. The intended condition should be stated more carefully, for example as Z_i depending only on the selected X_j's through some joint function.
  5. [Section 5.3, Table 3] The claim that SDSP can be incorporated into "most existing MDR methods" is supported by only three backbones (MMOE, PLE, SAR-Net); the abstract and introduction phrase this more strongly than the evidence warrants.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the selection reward is validation AUC, the reported result is held-out test AUC, and self-citations are peripheral.

full rationale

The paper's derivation chain is self-contained. The prototype-based distance (Eqs. 6-8) is learned from model representations via a reconstruction loss, and the dynamic selection (Algorithm 1) explicitly uses validation-domain metrics as the bandit reward ("Calculate metric {r_d} with validation data V"), while all reported results are measured on a separately held-out test set. The same metric name 'AUC' appears in both places, but the selection and evaluation sets are disjoint, so the reported gains are not the training reward by construction. The paper also discloses that selection uses "the supervised signals of the domain metrics and the unsupervised distance measure", so the unsupervised distance is not silently claimed to be the sole predictor. The ablation in Figure 4, where even a poor distance ("Average") outperforms "Random", is consistent with this disclosed design: the bandit's validation-AUC reward is intended to guide selection, so a poor distance still benefits from this supervision; this is a stated component, not a hidden equivalence. Remark 2's prefix-only restriction is an unproven search-space assumption, not a circular definition: it limits which states the bandit may visit but does not define the target result into existence. The self-citations in the paper (general recommendation references, prototype-learning citations, and an AUC-significance threshold) are peripheral and do not carry the core derivation. No load-bearing step reduces to its own inputs.

Assumptions & free parameters 8 free parameters · 6 assumptions · 2 invented entities

The central claim rests on several fitted hyperparameters, an unproven nested-prefix structure for the search space, and a strong optimality assumption in the theoretical analysis. The prototype and the asymmetric distance are internal constructs with no external falsifiable predictions. The masks selected by validation AUC are themselves fitted choices that directly shape the reported results.

free parameters (8)
  • Prototype learning coefficient gamma = 0.0001 (Movielens/Amazon), 0.001 (Douban)
    Searched over [1e-3, 1e-4, 1e-5]; controls reconstruction loss weight in Eq. (12).
  • Number of prototypes M = 10
    Chosen from Figure 3 hyperparameter analysis; best around 10.
  • Epsilon-greedy decay rate d_r = 0.9
    Set by hand in Section 5.1.4; controls exploration-exploitation trade-off.
  • Selection interval l = 2 iterations
    Algorithm 1 line 4; domain selection runs every 2 iterations, chosen without reported tuning.
  • Batch size = 4096
    Selected from Figure 3; larger batches improve stability of prototype learning.
  • Learning rate = 0.01 (Movielens), 0.0001 (Amazon), 0.001 (Douban)
    Searched over [0.01, 0.001, 0.0001]; Table 5.
  • Per-domain expert count E_d = 1
    Fixed for all domains; not tuned in main experiments.
  • Selected domain subsets S_d (mask) = Not reported per dataset; learned during training
    The masks are chosen via epsilon-greedy using validation AUC; they are fitted choices that directly determine the reported test performance.
assumptions (6)
  • domain assumption Optimality assumption: i = argmax_j I(X_j; Y_i) (Assumption 1)
    Used in Theorem 3 proof to set max_j I(X_j;Y_i) = I(X_i;Y_i); strong and untested, can fail in cold-start or sparse domains.
  • ad hoc to paper Candidate sets are prefixes of the distance-sorted list (Remark 2)
    Reduces search from O(2^(D-1)) to O(D) without proof that beneficial sets are nested; asymmetric transfer in Section 1 suggests non-prefix sets may be optimal.
  • domain assumption Prototype reconstruction loss ensures prototypes are representative (Eq. 8)
    Assumes that decoding ability of a few prototypes implies the prototypes capture transferable domain structure.
  • domain assumption Distance between prototypes reflects positive or negative transfer benefit (Definitions 1-2)
    No formal link is established between prototype L2 distance and mutual information or task transfer; Figure 6 gives only visual evidence.
  • standard math Standard information-theoretic identities: cross-entropy variational form, data-processing inequality (Lemma 4)
    Used in the proof of Theorem 3; these are standard.
  • domain assumption (X,Y) follows joint distribution p
    Assumed for theorem analysis; common but unverified for recommendation data.
invented entities (2)
  • Domain-level prototype p_d (M prototypes per domain)
    purpose: Represent each domain compactly so domain-to-domain distance can be computed in O(BM+M^2) instead of O(B^2); used to order candidate domain subsets.
    The prototype is an internal latent construct learned by an encoder-decoder with reconstruction loss on the model's own embeddings; its validity is only indirectly evaluated via downstream AUC, with no external falsifiable prediction.
  • Asymmetric domain distance dis(d1,d2) (Definition 2)
    purpose: Quantify how far domain d2 is from domain d1 to guide which source domains to include; asymmetry is claimed to model non-symmetric transfer relationships.
    A defined mathematical quantity, not an empirical observable; its correlation with actual transfer is only shown qualitatively in Figure 6.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Measure Domain's Gap: A Similar Domain Selection Principle for Multi-Domain Recommendation." pith.science (2026). https://pith.science/paper/WK7UR7T3

@misc{pith2026250520227,
  author       = {Pith},
  title        = {Pith review of: Measure Domain's Gap: A Similar Domain Selection Principle for Multi-Domain Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WK7UR7T3}},
  note         = {Machine review of arXiv:2505.20227}
}
read the original abstract

Multi-Domain Recommendation (MDR) achieves the desirable recommendation performance by effectively utilizing the transfer information across different domains. Despite the great success, most existing MDR methods adopt a single structure to transfer complex domain-shared knowledge. However, the beneficial transferring information should vary across different domains. When there is knowledge conflict between domains or a domain is of poor quality, unselectively leveraging information from all domains will lead to a serious Negative Transfer Problem (NTP). Therefore, how to effectively model the complex transfer relationships between domains to avoid NTP is still a direction worth exploring. To address these issues, we propose a simple and dynamic Similar Domain Selection Principle (SDSP) for multi-domain recommendation in this paper. SDSP presents the initial exploration of selecting suitable domain knowledge for each domain to alleviate NTP. Specifically, we propose a novel prototype-based domain distance measure to effectively model the complexity relationship between domains. Thereafter, the proposed SDSP can dynamically find similar domains for each domain based on the supervised signals of the domain metrics and the unsupervised distance measure from the learned domain prototype. We emphasize that SDSP is a lightweight method that can be incorporated with existing MDR methods for better performance while not introducing excessive time overheads. To the best of our knowledge, it is the first solution that can explicitly measure domain-level gaps and dynamically select appropriate domains in the MDR field. Extensive experiments on three datasets demonstrate the effectiveness of our proposed method.

Figures

Figures reproduced from arXiv: 2505.20227 by the authors.

Figure 1
Figure 1. The performance of STAR and M2M with selected [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The framework of the proposed SDSP. 3.3 Prototype-based Domain Distance Measure To measure the distance between domains and consider the learn￾ing ability of the model, an intuitive approach is to compute the similarity between domain-specific representations, i.e., calculating 𝑑𝑖𝑠(𝒉 𝑑 , 𝒉 𝑑 ′ ), where 𝒉 𝑑 is the learned representation in 𝑑-th domain. However, such an approach would incur a huge time overhead of O (… view at source ↗
Figure 3
Figure 3. Hyperparameter Analysis of MMOE+SDSP and PLE+SDSP on Movielens dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: The comparison of the negative transfer of PLE and [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: In the left figure, the horizontal axis indicates the do [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Intra-view and Inter-view Correlation Guided Multi-view Novel Class Discovery

    cs.CV 2025-07 conditional novelty 5.0 of 10

    A matrix-factorization method that learns view weights from labeled known classes and transfers them to cluster novel classes in multi-view data.

Reference graph

Works this paper leans on

83 extracted references · 36 canonical work pages · cited by 1 Pith paper

  1. [1]

    Guangji Bai and Liang Zhao. 2022. Saliency-regularized deep multi-task learning. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 15–25

    2022

  2. [2]

    Philippe C Besse, Brendan Guillouet, Jean-Michel Loubes, and François Royer

  3. [3]

    Weijie Bian, Kailun Wu, Lejian Ren, Qi Pi, Yujing Zhang, Can Xiao, Xiang-Rong Sheng, Yong-Nan Zhu, Zhangming Chan, Na Mou, et al. 2020. Can: Feature co- action for click-through rate prediction.arXiv preprint arXiv:2011.05625(2020)

    arXiv 2020

  4. [4]

    Rich Caruana. 1997. Multitask learning.Machine learning28 (1997), 41–75

    1997

  5. [5]

    Jianxin Chang, Chenbin Zhang, Yiqun Hui, Dewei Leng, Yanan Niu, Yang Song, and Kun Gai. 2023. Pepnet: Parameter and embedding personalized network for infusing with personalized prior information. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3795–3804

    2023

  6. [6]

    Lei Chen, Fajie Yuan, Jiaxi Yang, Xiangnan He, Chengming Li, and Min Yang

  7. [7]

    Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen Anderson, Greg Corrado, Wei Chai, Mustafa Ispir, et al

  8. [8]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep neural networks for youtube recommendations. InProceedings of the 10th ACM conference on recommender systems. 191–198

    2016

Show all 83 references
  1. [9]

    Wenqi Fan, Xiangyu Zhao, Qing Li, Tyler Derr, Yao Ma, Hui Liu, Jianping Wang, and Jiliang Tang. 2023. Adversarial attacks for black-box recommender sys- tems via copying transferable cross-domain user profiles.IEEE Transactions on Knowledge and Data Engineering35, 12 (2023), 1...

    2023

  2. [10]

    InProceedings of the 1st workshop on deep learning for recommender systems

    Wide & deep learning for recommender systems. InProceedings of the 1st workshop on deep learning for recommender systems. 7–10

  3. [11]

    Chunjing Gan, Bo Huang, Binbin Hu, Jian Ma, Zhiqiang Zhang, Jun Zhou, Guan- nan Zhang, and Wenliang Zhong. 2024. PEACE: Prototype lEarning Augmented transferable framework for Cross-domain rEcommendation. InProceedings of the 17th ACM International Conference on Web Search and...

    2024

  4. [12]

    Jingtong Gao, Bo Chen, Menghui Zhu, Xiangyu Zhao, Xiaopeng Li, Yuhao Wang, Yichao Wang, Huifeng Guo, and Ruiming Tang. 2024. HierRec: Scenario-Aware Hierarchical Modeling for Multi-scenario Recommendations. InProceedings of the 33rd ACM International Conference on Information ...

    2024

  5. [13]

    Zichuan Fu, Xiangyang Li, Chuhan Wu, Yichao Wang, Kuicai Dong, Xiangyu Zhao, Mengchen Zhao, Huifeng Guo, and Ruiming Tang. 2023. A unified frame- work for multi-domain ctr prediction via large language models.ACM Transac- tions on Information Systems(2023)

    2023

  6. [14]

    Yulong Gu, Wentian Bao, Dan Ou, Xiang Li, Baoliang Cui, Biyu Ma, Haikuan Huang, Qingwen Liu, and Xiaoyi Zeng. 2021. Self-supervised learning on users’ spontaneous behaviors for multi-scenario ranking in e-commerce. InProceedings of the 30th ACM International Conference on Info...

    2021

  7. [15]

    Renchu Guan, Haoyu Pang, Fausto Giunchiglia, Yanchun Liang, and Xiaoyue Feng. 2022. Cross-domain meta-learner for cold-start recommendation.IEEE Transactions on Knowledge and Data Engineering35, 8 (2022), 7829–7843

    2022

  8. [16]

    Jingtong Gao, Xiangyu Zhao, Bo Chen, Fan Yan, Huifeng Guo, and Ruiming Tang. 2023. AutoTransfer: Instance transfer for cross-domain recommendations. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1478–1487

    2023

  9. [17]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. InProceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval. 639–648

    2020

  10. [18]

    Yifei He, Runxiang Cheng, Gargi Balasubramaniam, Yao-Hung Hubert Tsai, and Han Zhao. 2024. Efficient Modality Selection in Multimodal Learning.Journal of Machine Learning Research25, 47 (2024), 1–39

    2024

  11. [19]

    Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: a factorization-machine based neural network for CTR prediction. In Proceedings of the 26th International Joint Conference on Artificial Intelligence. 1725–1731

    2017

  12. [20]

    Chi, Julian McAuley, and Derek Zhiyuan Cheng

    Yupeng Hou, Jianmo Ni, Zhankui He, Noveen Sachdeva, Wang-Cheng Kang, Ed H. Chi, Julian McAuley, and Derek Zhiyuan Cheng. 2025. ActionPiece: Contextually Tokenizing Action Sequences for Generative Recommendation. InICML

    2025

  13. [21]

    Jiarui Jin, Xianyu Chen, Weinan Zhang, Yuanbo Chen, Zaifan Jiang, Zekun Zhu, Zhewen Su, and Yong Yu. 2022. Multi-scale user behavior network for entire space multi-task learning. InProceedings of the 31st ACM International Conference on Information & Knowledge Management. 874–883

    2022

  14. [22]

    Yupeng Hou, Jiacheng Li, Ashley Shin, Jinsung Jeon, Abhishek Santhanam, Wei Shao, Kaveh Hassani, Ning Yao, and Julian McAuley. 2025. Generating Long Semantic IDs in Parallel for Recommendation. InKDD

    2025

  15. [23]

    Hanyu Li, Weizhi Ma, Peijie Sun, Jiayu Li, Cunxiang Yin, Yancheng He, Guoqiang Xu, Min Zhang, and Shaoping Ma. 2024. Aiming at the target: Filter collabora- tive information for cross-domain recommendation. InProceedings of the 47th International ACM SIGIR Conference on Resear...

    2024

  16. [24]

    Jinyun Li, Huiwen Zheng, Yuanlin Liu, Minfang Lu, Lixia Wu, and Haoyuan Hu

  17. [25]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In2018 IEEE international conference on data mining (ICDM). IEEE, 197–206

    2018

  18. [26]

    Junnan Li, Pan Zhou, Caiming Xiong, and Steven C. H. Hoi. 2021. Prototypi- cal Contrastive Learning of Unsupervised Representations. In9th International Conference on Learning Representations

    2021

  19. [27]

    Xiaopeng Li, Jingtong Gao, Pengyue Jia, Yichao Wang, Wanyu Wang, Yejing Wang, Yuhao Wang, Huifeng Guo, and Ruiming Tang. 2024. Scenario-Wise Rec: A Multi-Scenario Recommendation Benchmark.arXiv preprint arXiv:2412.17374 (2024)

    2024

  20. [28]

    Xinhang Li, Zhaopeng Qiu, Xiangyu Zhao, Zihao Wang, Yong Zhang, Chunxiao Xing, and Xian Wu. 2022. Gromov-wasserstein guided representation learning for cross-domain recommendation. InProceedings of the 31st ACM International Conference on Information & Knowledge Management. 1199–1208

    2022

  21. [29]

    Jinyun Li, Huiwen Zheng, Yuanlin Liu, Minfang Lu, Lixia Wu, and Haoyuan Hu. 2023. ADL: Adaptive Distribution Learning Framework for Multi-Scenario CTR Prediction. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval....

    2023

  22. [30]

    Fanghui Liu, Luca Viano, and Volkan Cevher. 2022. Understanding deep neural function approximation in reinforcement learning via 𝑒𝑝𝑠𝑖𝑙𝑜𝑛 -greedy explo- ration.Advances in Neural Information Processing Systems35 (2022), 5093–5108

    2022

  23. [31]

    Qidong Liu, Xian Wu, Yejing Wang, Zijian Zhang, Feng Tian, Yefeng Zheng, and Xiangyu Zhao. 2024. Llm-esr: Large language models enhancement for long- tailed sequential recommendation.Advances in Neural Information Processing Systems37 (2024), 26701–26727

    2024

  24. [32]

    Qidong Liu, Fan Yan, Xiangyu Zhao, Zhaocheng Du, Huifeng Guo, Ruiming Tang, and Feng Tian. 2023. Diffusion augmentation for sequential recommenda- tion. InProceedings of the 32nd ACM International conference on information and knowledge management. 1576–1586

    2023

  25. [33]

    Xiaopeng Li, Fan Yan, Xiangyu Zhao, Yichao Wang, Bo Chen, Huifeng Guo, and Ruiming Tang. 2023. Hamur: Hyper adapter for multi-domain recommendation. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management. 1268–1277

    2023

  26. [34]

    Yue Liu, Shihao Zhu, Jun Xia, Yingwei Ma, Jian Ma, Wenliang Zhong, Xinwang Liu, Shengju Yu, and Kejun Zhang. 2024. End-to-end Learnable Clustering for Intent Learning in Recommendation. InProc. of NeurIPS

    2024

  27. [35]

    Yue Liu, Shihao Zhu, Tianyuan Yang, Jian Ma, and Wenliang Zhong. 2024. Identify Then Recommend: Towards Unsupervised Group Recommendation. InProc. of NeurIPS

    2024

  28. [36]

    Huishi Luo, Yiwen Chen, Yiqing Wu, Fuzhen Zhuang, and Deqing Wang. 2024. One for Dozens: Adaptive REcommendation for All Domains with Counterfactual Augmentation.arXiv preprint arXiv:2412.11905(2024)

    2024 arXiv

  29. [37]

    Qi Liu, Zhilong Zhou, Gangwei Jiang, Tiezheng Ge, and Defu Lian. 2023. Deep task-specific bottom representation network for multi-task recommendation. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management. 1637–1646

    2023

  30. [38]

    Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H Chi. 2018. Modeling task relationships in multi-task learning with multi-gate mixture-of- experts. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 1930–1939

    2018

  31. [39]

    Xiao Ma, Liqin Zhao, Guan Huang, Zhi Wang, Zelin Hu, Xiaoqiang Zhu, and Kun Gai. 2018. Entire space multi-task model: An effective approach for estimating post-click conversion rate. InThe 41st International ACM SIGIR Conference on Research & Development in Information Retriev...

    2018

  32. [40]

    Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural...

    2019

  33. [41]

    Linhao Luo, Yumeng Li, Buyu Gao, Shuai Tang, Sinan Wang, Jiancheng Li, Tan- chao Zhu, Jiancai Liu, Zhao Li, and Shirui Pan. 2023. MAMDR: A model agnostic learning framework for multi-domain recommendation. In2023 IEEE 39th Inter- national Conference on Data Engineering (ICDE)....

    2023

  34. [42]

    Chung Park, Taesan Kim, Hyungjun Yoon, Junui Hong, Yelim Yu, Mincheol Cho, Minsung Choi, and Jaegul Choo. 2024. Pacer and Runner: Cooperative Learning Framework between Single-and Cross-Domain Sequential Recommendation. InProceedings of the 47th International ACM SIGIR Confere...

    2024

  35. [43]

    J Ben Schafer, Joseph A Konstan, and John Riedl. 2001. E-commerce recommen- dation applications.Data mining and knowledge discovery5 (2001), 115–153

    2001

  36. [44]

    Qijie Shen, Wanjie Tao, Jing Zhang, Hong Wen, Zulong Chen, and Quan Lu. 2021. SAR-Net: A scenario-aware ranking network for personalized fair recommenda- tion in hundreds of travel scenarios. InProceedings of the 30th ACM International Conference on Information & Knowledge Man...

    2021

  37. [45]

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

  38. [46]

    Weiping Song, Chence Shi, Zhiping Xiao, Zhijian Duan, Yewen Xu, Ming Zhang, and Jian Tang. 2019. Autoint: Automatic feature interaction learning via self- attentive neural networks. InProceedings of the 28th ACM international conference on information and knowledge management....

    2019

  39. [47]

    Trevor Standley, Amir Zamir, Dawn Chen, Leonidas Guibas, Jitendra Malik, and Silvio Savarese. 2020. Which tasks should be learned together in multi-task learning?. InInternational conference on machine learning. PMLR, 9120–9132

    2020

  40. [48]

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang

  41. [49]

    Tianxiang Sun, Yunfan Shao, Xiaonan Li, Pengfei Liu, Hang Yan, Xipeng Qiu, and Xuanjing Huang. 2020. Learning sparse sharing architectures for multiple tasks. InProceedings of the AAAI conference on artificial intelligence, Vol. 34. 8936–8943

    2020

  42. [50]

    Xiang-Rong Sheng, Liqin Zhao, Guorui Zhou, Xinyao Ding, Binding Dai, Qiang Luo, Siran Yang, Jingshan Lv, Chi Zhang, Hongbo Deng, et al. 2021. One model to serve all: Star topology adaptive recommender for multi-domain ctr prediction. In Proceedings of the 30th ACM Internationa...

    2021

  43. [51]

    Junxiong Tong, Mingjia Yin, Hao Wang, Qiushi Pan, Defu Lian, and Enhong Chen

  44. [52]

    María Cora Urdaneta-Ponte, Amaia Mendez-Zorrilla, and Ibon Oleagordia-Ruiz

  45. [53]

    Ajay Krishna Vajjala, Arun Krishna Vajjala, Ziwei Zhu, and David S Rosen- blum. 2024. Analyzing the Impact of Domain Similarity: A New Perspective in Cross-Domain Recommendation. In2024 International Joint Conference on Neural Networks (IJCNN). IEEE, 1–8

    2024

  46. [54]

    Xinhang Wan, Jiyuan Liu, Xinbiao Gan, Xinwang Liu, Siwei Wang, Yi Wen, Tianjiao Wan, and En Zhu. 2024. One-Step Multi-View Clustering With Diverse Representation.IEEE Transactions on Neural Networks and Learning Systems (2024), 1–13. https://doi.org/10.1109/TNNLS.2024.3378194

    2024

  47. [55]

    Ruoxi Wang, Bin Fu, Gang Fu, and Mingliang Wang. 2017. Deep & cross network for ad click predictions. InProceedings of the ADKDD’17. 1–7

    2017

  48. [56]

    Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. 2020. Progressive layered extraction (ple): A novel multi-task learning (mtl) model for personalized recommendations. InProceedings of the 14th ACM Conference on Recommender Systems. 269–278

    2020

  49. [57]

    Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural graph collaborative filtering. InProceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval. 165–174

    2019

  50. [58]

    Yuhao Wang, Ziru Liu, Yichao Wang, Xiangyu Zhao, Bo Chen, Huifeng Guo, and Ruiming Tang. 2024. Diff-MSR: A diffusion model enhanced paradigm for cold-start multi-scenario recommendation. InProceedings of the 17th ACM International Conference on Web Search and Data Mining. 779–787

    2024

  51. [59]

    Yuhao Wang, Xiangyu Zhao, Bo Chen, Qidong Liu, Huifeng Guo, Huanshuo Liu, Yichao Wang, Rui Zhang, and Ruiming Tang. 2023. PLATE: A prompt-enhanced paradigm for multi-scenario recommendations. InProceedings of the 46th In- ternational ACM SIGIR Conference on Research and Develo...

    2023

  52. [60]

    Recommendation systems for education: Systematic review.Electronics10, 14 (2021), 1611

    2021

  53. [61]

    Bencheng Yan, Pengjie Wang, Kai Zhang, Feng Li, Hongbo Deng, Jian Xu, and Bo Zheng. 2022. Apg: Adaptive parameter generation network for click-through rate prediction.Advances in Neural Information Processing Systems35 (2022), 24740–24752

    2022

  54. [62]

    Xuanhua Yang, Xiaoyu Peng, Penghui Wei, Shaoguo Liu, Liang Wang, and Bo Zheng. 2022. Adasparse: Learning adaptively sparse structures for multi-domain click-through rate prediction. InProceedings of the 31st ACM International Con- ference on Information & Knowledge Management....

    2022

  55. [63]

    Zhengyi Yang, Jiancan Wu, Zhicai Wang, Xiang Wang, Yancheng Yuan, and Xiangnan He. 2024. Generate what you prefer: Reshaping sequential recommen- dation via guided diffusion.Advances in Neural Information Processing Systems 36 (2024)

    2024

  56. [64]

    Xiang Wang, Xiangnan He, Liqiang Nie, and Tat-Seng Chua. 2017. Item silk road: Recommending items from information domains to social users. InProceedings of the 40th International ACM SIGIR conference on Research and Development in Information Retrieval. 185–194

    2017

  57. [65]

    Qianqian Zhang, Xinru Liao, Quan Liu, Jian Xu, and Bo Zheng. 2022. Leaving no one behind: A multi-scenario multi-task meta learning approach for advertiser modeling. InProceedings of the Fifteenth ACM International Conference on Web Search and Data Mining. 1368–1376

    2022

  58. [66]

    Wen Zhang, Lingfei Deng, Lei Zhang, and Dongrui Wu. 2022. A survey on negative transfer.IEEE/CAA Journal of Automatica Sinica10, 2 (2022), 305–329

    2022

  59. [67]

    Xiangyu Zhao, Long Xia, Liang Zhang, Zhuoye Ding, Dawei Yin, and Jiliang Tang. 2018. Deep reinforcement learning for page-wise recommendations. In Proceedings of the 12th ACM conference on recommender systems. 95–103

    2018

  60. [68]

    Lianghao Xia, Chao Huang, Yong Xu, Peng Dai, Xiyue Zhang, Hongsheng Yang, Jian Pei, and Liefeng Bo. 2021. Knowledge-enhanced hierarchical graph trans- former network for multi-behavior recommendation. InProceedings of the AAAI conference on artificial intelligence, Vol. 35. 4486–4493

    2021

  61. [69]

    Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep interest network for click-through rate prediction. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. ...

    2018

  62. [70]

    Kun Zhou, Hui Yu, Wayne Xin Zhao, and Ji-Rong Wen. 2022. Filter-enhanced MLP is all you need for sequential recommendation. InProceedings of the ACM web conference 2022. 2388–2399

    2022

  63. [71]

    Chenxu Zhu, Peng Du, Xianghui Zhu, Weinan Zhang, Yong Yu, and Yang Cao

  64. [72]

    Shengju Yu, Siwei Wang, Yi Wen, Ziming Wang, Zhigang Luo, En Zhu, and Xinwang Liu. 2023. How to construct corresponding anchors for incomplete mul- tiview clustering.IEEE Transactions on Circuits and Systems for Video Technology 34, 4 (2023), 2845–2860

    2023

  65. [76]

    Xiangyu Zhao, Liang Zhang, Zhuoye Ding, Long Xia, Jiliang Tang, and Dawei Yin

  66. [81]

    InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    User-tag profile modeling in recommendation system via contrast weighted tag masking. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4630–4638

  67. [82]

    Jiachen Zhu, Yichao Wang, Jianghao Lin, Jiarui Qin, Ruiming Tang, Weinan Zhang, and Yong Yu. 2024. M-scan: A Multi-Scenario Causal-driven Adaptive Network for Recommendation. InProceedings of the ACM on Web Conference

    2024

  68. [83]

    3844–3853. A THEOREM PROOF To proof the Theorem 3, we first need proof the following lemma: Lemma 4.[18] For any domain subsetD⊆D, we have 𝐻(Y|X D)=inf 𝑓 E𝑝[ℓCE(𝑓(X D),Y)] Proof.Lemma1 E𝑝[ℓCE(𝑓(X D),Y)] =−E𝑝[I(Y=0)log(1−𝑓(X D))+I(Y=1)log(𝑓(X D))] =−E𝐷 E𝑌[I(Y=0)log(1−𝑓(X D))+I(...

    2025

  69. [2016]

    IEEE Transactions on Intelligent Transportation Systems17, 11 (2016), 3306–3317

    Review and perspective for distance-based clustering of vehicle trajectories. IEEE Transactions on Intelligent Transportation Systems17, 11 (2016), 3306–3317

    2016

  70. [2018]

    InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining

    Recommendations with negative feedback via pairwise deep reinforcement learning. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. 1040–1048

  71. [2019]

    InProceedings of the 28th ACM international conference on information and knowledge management

    BERT4Rec: Sequential recommendation with bidirectional encoder rep- resentations from transformer. InProceedings of the 28th ACM international conference on information and knowledge management. 1441–1450

  72. [2021]

    IEEE Transactions on Knowledge and Data Engineering35, 3 (2021), 3239–3252

    User-specific adaptive fine-tuning for cross-domain recommendations. IEEE Transactions on Knowledge and Data Engineering35, 3 (2021), 3239–3252

    2021

  73. [2022]

    AdaptDHM: Adaptive Distribution Hierarchical Model for Multi-Domain CTR Prediction.arXiv preprint arXiv:2211.12105(2022)

    2022 arXiv

  74. [2023]

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

    Multi-domain recommendation with embedding disentangling and domain alignment. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management. 1917–1927

    1917

  75. [2024]

    InInternational Conference on Web Information Systems Engineering

    MDAP: A Multi-view Disentangled and Adaptive Preference Learning Framework for Cross-Domain Recommendation. InInternational Conference on Web Information Systems Engineering. Springer, 164–178

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.