MoE-MLoRA for Multi-Domain CTR Prediction: Efficient Adaptation with Expert Specialization

Aviel Ben Siman Tov; Eyal German; Ken Yaggel

REVIEW 4 major objections 7 minor 23 references

MoE-MLoRA for Multi-Domain CTR Prediction: Efficient Adaptation with Expert Specialization

T0 review · 4 major / 7 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 MoE-MLoRA claims sparse multi-domain CTR gains of +1.45 WAUC over MLoRA.

desk verdict A clear, honest empirical sketch whose central gating claim is not yet demonstrated; the one-expert-per-domain setup makes the reported gains equally explainable as per-domain LoRA selection. read the letter →

arxiv 2506.07563 v3 pith:STVWV6P2 submitted 2025-06-09 cs.IR cs.AI

Ken Yaggel , Eyal German , Aviel Ben Siman Tov This is my paper

classification cs.IRcs.AI

keywords mixture-of-expertsmulti-domainCTRpredictionlow-rankadaptationLoRAgatingnetworkexpertspecializationparameter-efficientfine-tuningrecommendationsystems

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 tries to establish that replacing MLoRA's single per-domain LoRA adapter with a mixture of domain-specialized LoRA experts, combined with a learned gating network, improves multi-domain click-through-rate prediction in sparse, heterogeneous settings. On Taobao, split into 10 or 20 themes, the method beats MLoRA by average +1.43 and +1.45 weighted AUC; on the denser, less diverse Movielens splits, gains are small or slightly negative. If correct, this gives a practical three-stage recipe—pretrain the backbone, train each expert alone, then train the gate—for adapting recommendation models across many domains. A reader should care because multi-domain personalization is common in production, and the proposed change is parameter-efficient and applied across eight CTR model architectures.

What carries the argument

The carrying mechanism is the MloraMoE layer, in which each LoRA adapter is treated as an expert. Training proceeds in three phases: first the backbone model is pretrained on all domains, then each domain-specific LoRA expert is trained independently while the backbone and all other experts are frozen, and finally a gating network is trained to weight the frozen experts' outputs using the domain as an additional input. This separation is what lets each expert specialize without interference while the gate performs cross-domain knowledge transfer.

What would settle it

Re-run MoE-MLoRA and MLoRA on Taobao-20 under identical conditions across at least five random seeds; if the average WAUC gap falls within one standard deviation of zero, the claimed +1.45 improvement does not hold.

Watch

Extended reading notes

Core claim

The central claim is that a mixture-of-experts layer built from LoRA adapters, with one adapter per domain trained independently and then frozen, outperforms the single-adapter-per-domain MLoRA baseline when domains are sparse and behaviorally distinct. On the Taobao-10 and Taobao-20 theme splits, MoE-MLoRA improves average weighted AUC by +1.43 and +1.45 percentage points over MLoRA across eight models, while on Movielens gender and age splits it is slightly worse on average (−0.14 and −0.10) and only modestly better on the 21-occupation split (+0.11). The paper also claims that adding more experts per domain does not consistently help; the best ensemble size is model-dependent, so expert count should be tuned per architecture rather than increased blindly.

Load-bearing premise

The comparison assumes that the MLoRA baseline and domain splits are reproduced faithfully and that the reported single-run WAUC differences are not seed or tuning noise, since the paper reports no variance or significance tests.

Editorial extensions

If this is right

On sparse, theme-separated datasets like Taobao-20, switching MLoRA's single adapter to frozen per-domain experts plus a gate gives about 1.45 WAUC points on average across eight models.
On dense, low-diversity splits such as Movielens gender and age, the same setup is not reliably better and can lose a few tenths of a point, so the benefit appears only with meaningful domain separation.
Granular partitions that expose real behavioral differences, such as the 21-occupation Movielens split, can benefit modestly even in otherwise dense data.
More experts per domain does not guarantee better predictions; the best ensemble size varies by model, so expert count should be tuned per architecture.
The method remains parameter-efficient because only small LoRA adapters and the gating network are trained after the backbone is frozen.

Reading between the lines

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

If the gains survive replication with variance, the sequential freeze-then-gate schedule is likely the active ingredient; joint end-to-end MoE training may blur the specialization that drives the Taobao improvement.
The pattern suggests MoE-MLoRA would help most where domain boundaries align with real user-behavior differences and data is fragmented; a preprocessing step that estimates domain separability could predict when to apply it.
The lack of improvement on dense Movielens splits implies the gating function may need input features beyond domain identity to be useful when domains are similar.
A testable extension is to measure whether gate weights correlate with domain sparsity or distributional divergence; if they do not, the mechanism behind the Taobao gain remains unclear.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Short version: this is a clearly written, honest empirical paper, but the central claim — that a learned gate over frozen per-domain LoRA experts improves multi-domain CTR — is not actually demonstrated. With one expert per domain and the domain ID as gate input, the gate can collapse to a one-hot selector. Then the Taobao gains are just per-domain LoRA fine-tuning with a domain switch, not adaptive mixing.

What is new: the specific three-phase scheme — pretrain backbone, fine-tune each expert independently and freeze, then train the gate — applied to eight CTR models on Taobao and Movielens. The two prior MoE-LoRA papers train experts jointly, so the sequential training is a legitimate extension. Credit where due: the paper reports negative Movielens results honestly, and the expert-count experiment gives a clear, model-dependent picture. That part would be useful to practitioners.

Soft spots, in proportion. The missing gate ablation is the big one. Section 4.1 only compares the full MoE-MLoRA against MLoRA; there is no variant that removes the gate, no analysis of learned routing weights or entropy, no hard-selection baseline. Since the primary configuration uses exactly one expert per domain and the gate receives the domain as input, the trivial solution is to select the domain's own expert. The stress-test is correct: nothing in the paper shows the gate is doing work beyond selection. This is a load-bearing gap, not a minor omission.

Second, Table 1 has a baseline-integrity issue. Several MLoRA cells are well below their own Base — AutoInt on Taobao-10 (72.09 vs 74.19) and DCN on Taobao-20 (74.02 vs 76.54). That suggests the baseline was undertuned and inflates the reported deltas. Combined with the lack of repeated seeds, significance tests, and hyperparameters, the +1.4 WAUC claims are plausible but not established. The abstract mentions a GitHub link but the paper body gives no URL, which is annoying but fixable.

On the upside, the comparison is not circular: the metric is held-out WAUC, the baseline is external prior work, and the authors do not overclaim. They explicitly say gains are dataset-dependent. That is more honest than much of the applied literature.

Who is this for: practitioners in multi-domain recommendation who want a quick, readable sense of what three-stage LoRA-MoE might do, and a warning that expert count needs model-aware tuning. As submitted, the empirical evidence is too thin to support the mechanism. I would not cite it yet. But it deserves serious review if the authors add a no-gate per-domain LoRA baseline, analyze gate weights, add repeated seeds, and tune the baselines. A workshop-style major revision would be appropriate.

Referee Report

4 major / 7 minor

Summary. The paper proposes MoE-MLoRA, a three-stage approach for multi-domain CTR prediction. A backbone model is first pretrained; per-domain LoRA experts are then fine-tuned independently with the rest of the model frozen; finally, a gating network is trained to combine the frozen experts, using the domain as an input. The method is evaluated on eight CTR backbone architectures (WDL, NFM, AutoInt, PNN, DCN, FiBiNET, DeepFM, xDeepFM) on five domain splits (Movielens gender/age/occupation and Taobao-10/20 themes), always against an MLoRA baseline. Results are mixed: consistent WAUC gains on the two Taobao splits (+1.43 and +1.45 average), a small positive gain on Movielens occupation (+0.11), and negative deltas on Movielens gender (−0.14) and age (−0.10). A second experiment varies the number of experts per domain on Movielens gender and reports model-dependent behavior. The paper concludes that the method is most beneficial in sparse, heterogeneous, large-scale datasets.

Significance. If the claimed results hold, the paper provides a simple and practical recipe—frozen per-domain LoRA experts plus a learned gate—that improves CTR prediction in sparse, heterogeneous multi-domain settings while being computationally cheap. The systematic evaluation across eight backbone architectures and the honest reporting of negative results on dense, homogeneous splits are useful contributions. The finding that the optimal number of experts is model-dependent is also informative for practitioners. However, the significance is currently limited by the absence of variance estimates, an ablation of the gating mechanism, and a fair baseline comparison; these issues prevent the empirical claims from being fully load-bearing.

major comments (4)

[§4.1, Table 1] The evaluation compares only the full MoE-MLoRA pipeline with MLoRA; there is no ablation in which the gating network is removed or replaced by a hard selection rule. Since the primary configuration uses exactly one expert per domain (Section 4.2) and the gate receives the domain as an input (Section 3), the gate could collapse to a one-hot mapping, making the reported gains equivalent to per-domain LoRA fine-tuning plus selection. The paper does not report learned gate weights, routing entropy, or a hard-selection baseline, so the contribution of the 'adaptive mixing' component is not established.
[Table 1, Figure 2] All results appear to be single runs; no standard deviations, confidence intervals, or significance tests are reported. Many deltas are small (e.g., Movielens-occupation Avg(Δ)=+0.11, Movielens-age Avg(Δ)=−0.10, individual cells such as WDL gender +0.05), which are within typical AUC seed noise for CTR models. The headline Taobao-20 gain of +1.45 WAUC needs to be shown to be robust across random seeds and repeated training runs.
[§4.1, Table 1] Several MLoRA results are below their own Base, e.g., Taobao-10 AutoInt 72.09 vs Base 74.19 and Taobao-20 DCN 74.02 vs Base 76.54. This suggests the MLoRA baseline may be undertuned, which would inflate the reported improvements. The claim that both models are trained under identical conditions does not address hyperparameter selection for the baseline; the paper should report the MLoRA configuration, whether it matches the original MLoRA paper, and show that the baseline tuning is fair.
[§5.1] The aggregate statistic Avg(Δ)=+0.70 is computed by averaging per-dataset averages, which gives equal weight to the two Taobao splits and the three Movielens splits. This obscures the fact that the method underperforms on Movielens gender (−0.14) and age (−0.10). The paper should present per-split results with uncertainty and avoid a single unweighted average that can be read as implying an overall improvement.

minor comments (7)

[Abstract, §1] The abstract and Section 1 state that code is available, but no GitHub URL or repository identifier appears anywhere in the manuscript.
[Eq. (1), Table 1] Equation (1) defines WAUC as a weighted average, but the abstract and Table 1 use the spelling 'Weighed-AUC'; the spelling should be consistent ('Weighted AUC').
[§2.2] The sentence 'In the MLoRA paper [21], they proposed' is awkward; rephrase, for example, 'MLoRA [21] proposed a technique for CTR prediction in a multi-domain setting'.
[Figure 2] Figure 2 lacks error bars or numerical values; since the differences between expert counts are small, a table with exact WAUC values would be more informative.
[Table 1 caption] The caption states 'one expert per layer for each domain', while Section 4.2 describes 'one expert per domain'; clarify whether experts are attached per layer or per model.
[References [7]] The Movielens dataset is cited as [7], but reference [7] (He and McAuley 2016) concerns fashion trend modeling, not the Movielens dataset; the canonical citation (Harper and Konstan, 'The MovieLens Datasets: History and Context') should be used.
[§1] The claim that the approach 'achieves state-of-the-art performance' is not supported by comparisons to other multi-domain methods such as PLE; either add such comparisons or soften the claim to 'competitive with MLoRA'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the claimed gains are empirical comparisons against an external baseline on held-out data.

full rationale

MoE-MLoRA is presented as an architecture plus a three-phase training procedure (Sections 3 and 4) and is evaluated by weighted AUC on Taobao and Movielens. The comparison baseline, MLoRA, is prior external work (reference [21]), and the datasets are public benchmarks with domain splits taken from that external paper. The reported +1.45 WAUC on Taobao-20 is a measured difference between two trained systems on test data, not a quantity that is defined in terms of the method's own outputs. The gate and experts are trained before evaluation, and WAUC is computed from held-out predictions via Eq. (1), so there is no fitted input being renamed as a prediction. Section 5.2's expert-count analysis is explicitly exploratory and does not define the main claim by construction. The absence of variance and significance tests, and the possibility that the MLoRA baseline is undertuned (e.g., several MLoRA cells below Base in Table 1), are empirical-validity concerns, not circularity. There is no self-citation chain, no uniqueness theorem, and no ansatz imported from the authors' prior work, so the derivation chain is self-contained as an empirical study.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central claim rests on one architectural invention, four unreported hyperparameter families, and four domain or training assumptions. The most consequential free parameter is the number of experts per domain, which the paper itself shows is model-dependent. The axioms are all plausible but none are tested against alternatives, so the ledger is dominated by design choices rather than derived quantities.

free parameters (4)

number_of_experts_per_domain = 1 in main experiments; varied 1-4 in Section 5.2
The main claim uses one expert per domain; Section 5.2 shows performance depends on this choice, so it is a tuned design parameter rather than a derived one.
LoRA rank = not reported
Rank controls adapter capacity; without it, the comparison to MLoRA is under-specified.
gating_network_hyperparameters = not reported
Layer sizes, input features, and regularization of the gate are not given; the gate is the component responsible for the claimed improvement.
training_phase_hyperparameters = not reported
Epochs, learning rates, and batch sizes for each of the three phases are absent, preventing replication.

assumptions (4)

domain assumption The domain splits create behaviorally distinct groups.
Section 4 splits Taobao by themes and Movielens by gender, age, and occupation; Section 6 itself concedes the gender and age splits may be too similar, so the method's success depends on this premise.
domain assumption The MLoRA baseline and hyperparameters are reproduced faithfully.
Section 4.1 states identical conditions but does not report the baseline's hyperparameters, tuning budget, or code version.
domain assumption Domain labels are available as input to the gate.
Section 3 says the gating function uses the domain as an additional input; in many production settings the domain may need to be inferred rather than observed.
ad hoc to paper Independent and frozen expert training is a sound optimization strategy.
Phases 2 and 3 fix experts before gating; no comparison against joint training is provided, so the benefit of this ordering is assumed.

invented entities (1)

MloraMoElayer (MoE-MLoRA module)
purpose: Replace a single MLoRA adapter with per-domain LoRA experts plus a gating function.
The module is defined only in this paper and validated only on the paper's own benchmarks; no external replication or theoretical analysis is provided.

how reviews work

0 comments

Cite this review

Pith. "Pith review of MoE-MLoRA for Multi-Domain CTR Prediction: Efficient Adaptation with Expert Specialization." pith.science (2026). https://pith.science/paper/STVWV6P2

@misc{pith2026250607563,
  author       = {Pith},
  title        = {Pith review of: MoE-MLoRA for Multi-Domain CTR Prediction: Efficient Adaptation with Expert Specialization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/STVWV6P2}},
  note         = {Machine review of arXiv:2506.07563}
}

read the original abstract

Personalized recommendation systems must adapt to user interactions across different domains. Traditional approaches like MLoRA apply a single adaptation per domain but lack flexibility in handling diverse user behaviors. To address this, we propose MoE-MLoRA, a mixture-of-experts framework where each expert is first trained independently to specialize in its domain before a gating network is trained to weight their contributions dynamically. We evaluate MoE-MLoRA across eight CTR models on Movielens and Taobao, showing that it improves performance in large-scale, dynamic datasets (+1.45 Weighed-AUC in Taobao-20) but offers limited benefits in structured datasets with low domain diversity and sparsity. Further analysis of the number of experts per domain reveals that larger ensembles do not always improve performance, indicating the need for model-aware tuning. Our findings highlight the potential of expert-based architectures for multi-domain recommendation systems, demonstrating that task-aware specialization and adaptive gating can enhance predictive accuracy in complex environments. The implementation and code are available in our GitHub repository.

Figures

Figures reproduced from arXiv: 2506.07563 by the authors.

**Figure 1.** Overview of the structured three-phase training process. The workflow illustrates the integration of MoE with LoRA, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Effect of the number of experts on W-AUC perfor [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

23 extracted references · 8 canonical work pages

[1]

Weijie Bian, Kailun Wu, Lejian Ren, Qi Pi, Yujing Zhang, Can Xiao, Xiang- Rong Sheng, Yong-Nan Zhu, Zhangming Chan, Na Mou, Xinchen Luo, Shiming Xiang, Guorui Zhou, Xiaoqiang Zhu, and Hongbo Deng. 2021. CAN: Feature Co-Action for Click-Through Rate Prediction. arXiv:2011.05625 [cs.IR] https: //arxiv.org/abs/2011.05625

arXiv 2021
[2]

Mathieu Blondel, Akinori Fujino, Naonori Ueda, and Masakazu Ishihata. 2016. Higher-order factorization machines. InProceedings of the 30th International Conference on Neural Information Processing Systems(Barcelona, Spain)(NIPS’16). Curran Associates Inc., Red Hook, NY, USA, 3359–3367

work page 2016
[3]

Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen Anderson, Greg Corrado, Wei Chai, Mustafa Ispir, Rohan Anil, Zakaria Haque, Lichan Hong, Vihan Jain, Xiaobing Liu, and Hemal Shah
[4]

Mark Dredze, Alex Kulesza, and Koby Crammer. 2010. Multi-domain learning by confidence-weighted parameter combination.Machine Learning79 (2010), 123–149

work page 2010
[5]

Zhengxiao Du, Xiaowei Wang, Hongxia Yang, Jingren Zhou, and Jie Tang. 2019. Sequential scenario-specific meta learner for online recommendation. InProceed- ings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2895–2904

work page 2019
[6]

Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: A Factorization-Machine based Neural Network for CTR Prediction. arXiv:1703.04247 [cs.IR] https://arxiv.org/abs/1703.04247

arXiv 2017
[7]

Ruining He and Julian McAuley. 2016. Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering. Inproceedings of the 25th international conference on world wide web. 507–517

2016
[8]

Xiangnan He and Tat-Seng Chua. 2017. Neural Factorization Machines for Sparse Predictive Analytics. InProceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval. 355–364

2017

Show all 23 references

[9]

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. 2022. Lora: Low-rank adaptation of large language models.ICLR1, 2 (2022), 3

2022
[10]

Tao Huang, Zhiqiang Zhang, Junlin Zhang, Yongjun Zhu, Rui Xu, Jiang Bian, Rui Xie, and Ruoming Jin. 2019. FiBiNET: Combining Feature Importance and Bilinear Feature Interaction for Click-Through Rate Prediction. InProceedings of the 13th ACM Conference on Recommender Systems. 169–177

2019
[11]

Yuchin Juan, Yong Zhuang, Wei-Sheng Chin, and Chih-Jen Lin. 2016. Field- aware Factorization Machines for CTR Prediction. InProceedings of the 10th ACM Conference on Recommender Systems(Boston, Massachusetts, USA)(RecSys ’16). Association for Computing Machinery, New York, NY,...

2016
[12]

Hospedales

Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. 2017. Deeper, Broader and Artier Domain Generalization. InProceedings of the IEEE Interna- tional Conference on Computer Vision (ICCV)

2017
[13]

Pengcheng Li, Runze Li, Qing Da, An-Xiang Zeng, and Lijun Zhang. 2020. Improv- ing multi-scenario learning to rank in e-commerce by exploiting task relation- ships in the label space. InProceedings of the 29th ACM International Conference on Information & Knowledge Management....

2020
[14]

Jianxun Lian, Xiaohuan Zhou, Fuzheng Zhang, Zhongxia Chen, Xing Xie, and Guangzhong Sun. 2018. xDeepFM: Combining Explicit and Implicit Feature Interactions for Recommender Systems.arXiv preprint arXiv:1803.05170(2018)

2018 arXiv
[15]

Qidong Liu, Xian Wu, Xiangyu Zhao, Yuanshao Zhu, Derong Xu, Feng Tian, and Yefeng Zheng. 2023. Moelora: An moe-based parameter efficient fine-tuning method for multi-task medical applications.CoRR(2023)

2023
[16]

Tongxu Luo, Jiahe Lei, Fangyu Lei, Weihao Liu, Shizhu He, Jun Zhao, and Kang Liu. 2024. Moelora: Contrastive learning guided mixture of experts on parameter- efficient fine-tuning for large language models.arXiv preprint arXiv:2402.12851 (2024)

2024 arXiv
[17]

Yanru Qu, Han Cai, Kan Ren, Weinan Zhang, Yong Yu, Ying Wen, and Jun Wang. 2016. Product-Based Neural Networks for User Response Prediction. In2016 IEEE 16th International Conference on Data Mining (ICDM). 1149–1154. https://doi.org/10.1109/ICDM.2016.0151

2016
[18]

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(B...

2019
[19]

Hongyan Tang, Junning Liu, Ming Zhao, and Xudong Gong. 2020. Progres- sive Layered Extraction (PLE): A Novel Multi-Task Learning (MTL) Model for Personalized Recommendations. InProceedings of the 14th ACM Conference on Recommender Systems(Virtual Event, Brazil)(RecSys ’20). As...

2020 doi
[20]

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
[21]

Zhiming Yang, Haining Gao, Dehong Gao, Luwei Yang, Libin Yang, Xiaoyan Cai, Wei Ning, and Guannan Zhang. 2024. MLoRA: Multi-Domain Low-Rank Adaptive Network for CTR Prediction. InProceedings of the 18th ACM Conference on Recommender Systems(Bari, Italy)(RecSys ’24). Associatio...

2024
[22]

Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. 2023. Domain Generalization: A Survey.IEEE Transactions on Pattern Analysis and Machine Intelligence45, 4 (2023), 4396–4415. https://doi.org/10.1109/TPAMI. 2022.3195549

2023
[2016]

InProceedings of the 1st Workshop on Deep Learning for Recommender Systems(Boston, MA, USA) (DLRS 2016)

Wide & Deep Learning for Recommender Systems. InProceedings of the 1st Workshop on Deep Learning for Recommender Systems(Boston, MA, USA) (DLRS 2016). Association for Computing Machinery, New York, NY, USA, 7–10. https://doi.org/10.1145/2988450.2988454

2016

Pith tools

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

[1] [1]

Weijie Bian, Kailun Wu, Lejian Ren, Qi Pi, Yujing Zhang, Can Xiao, Xiang- Rong Sheng, Yong-Nan Zhu, Zhangming Chan, Na Mou, Xinchen Luo, Shiming Xiang, Guorui Zhou, Xiaoqiang Zhu, and Hongbo Deng. 2021. CAN: Feature Co-Action for Click-Through Rate Prediction. arXiv:2011.05625 [cs.IR] https: //arxiv.org/abs/2011.05625

arXiv 2021

[2] [2]

Mathieu Blondel, Akinori Fujino, Naonori Ueda, and Masakazu Ishihata. 2016. Higher-order factorization machines. InProceedings of the 30th International Conference on Neural Information Processing Systems(Barcelona, Spain)(NIPS’16). Curran Associates Inc., Red Hook, NY, USA, 3359–3367

work page 2016

[3] [3]

Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen Anderson, Greg Corrado, Wei Chai, Mustafa Ispir, Rohan Anil, Zakaria Haque, Lichan Hong, Vihan Jain, Xiaobing Liu, and Hemal Shah

[4] [4]

Mark Dredze, Alex Kulesza, and Koby Crammer. 2010. Multi-domain learning by confidence-weighted parameter combination.Machine Learning79 (2010), 123–149

work page 2010

[5] [5]

Zhengxiao Du, Xiaowei Wang, Hongxia Yang, Jingren Zhou, and Jie Tang. 2019. Sequential scenario-specific meta learner for online recommendation. InProceed- ings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2895–2904

work page 2019

[6] [6]

Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: A Factorization-Machine based Neural Network for CTR Prediction. arXiv:1703.04247 [cs.IR] https://arxiv.org/abs/1703.04247

arXiv 2017

[7] [7]

Ruining He and Julian McAuley. 2016. Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering. Inproceedings of the 25th international conference on world wide web. 507–517

2016

[8] [8]

Xiangnan He and Tat-Seng Chua. 2017. Neural Factorization Machines for Sparse Predictive Analytics. InProceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval. 355–364

2017