Pith. sign in

REVIEW 4 major objections 6 minor 6 cited by

MTGR: Industrial-Scale Generative Recommendation Framework in Meituan

T0 review · 4 major / 6 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 MTGR, a generative recommender built on HSTU, keeps the cross features of DLRM and reports 65x per-sample FLOPs with power-law scaling and production gains.

desk verdict Genuinely new industrial combination of HSTU and cross features; scaling-law and cross-feature-indispensability claims are thinner than the prose suggests. read the letter →

arxiv 2505.18654 v4 pith:LZX5FAUX submitted 2025-05-24 cs.IR

Ruidong Han , Bin Yin , Shangyu Chen , He Jiang , Fei Jiang , Xiang Li , Chi Ma , Mincong Huang
show 7 more authors
Xiaoguang Li Chunzhen Jing Yueming Han Menglei Zhou Lei Yu Chuan Liu Wei Lin
This is my paper · ORCID
classification cs.IR
keywords generativerecommendationscalinglawcrossfeaturesHSTUrankingmodelindustrialGroup-LayerNormalizationuser-levelaggregation
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 proposes MTGR, a production ranking model that brings transformer-based generative recommendation to a major food delivery platform while keeping the handcrafted cross features that classical DLRM models depend on. Its central claim is that dropping these cross features, as earlier generative recommenders do, severely hurts performance, and that no amount of added compute makes up for it. MTGR attaches each candidate's cross features to its token and scores all candidates for a user in one aggregated forward pass, so computation grows sublinearly with candidate count. The paper reports a 65x increase in per-sample forward FLOPs relative to the production DLRM, the largest offline and online gains in nearly two years, and a power-law relationship between model compute and ranking quality. If the claim is right, generative recommenders can scale without sacrificing the feature engineering that industrial systems rely on.

What carries the argument

The load-bearing mechanism is a user-level token sequence fed through stacked HSTU (Hierarchical Sequential Transduction Units) self-attention blocks. Each scalar feature, historical item, real-time interaction, and candidate becomes a token, with cross features concatenated directly into the candidate tokens. Group-Layer Normalization normalizes tokens by semantic group before attention so different feature spaces align; a customized dynamic mask lets static user features attend to all tokens, real-time interactions attend only causally to later tokens, and each candidate attend only to itself. This arrangement lets one user representation serve all candidates in a single forward pass, so inference cost grows sublinearly with candidate count while the cross features remain in the input and drive performance.

What would settle it

Train MTGR-large without cross features and compare it directly with the strongest DLRM baseline; if the no-cross-feature model matches or exceeds DLRM, or if its scaling curve crosses DLRM at larger compute, the central premise fails. The paper's own ablation only tests MTGR-small, so this experiment would settle the claim.

Watch

Extended reading notes

Core claim

MTGR's core discovery is that a generative ranking model can retain all DLRM features, including handcrafted cross features, within the HSTU architecture and still scale like a generative recommender. The model reorganizes each user's request into a token sequence: user profile features, historical and real-time behavior sequences, and candidate items where each candidate token carries its own cross features. An HSTU encoder with Group-Layer Normalization and a customized dynamic mask processes this sequence, and the candidate outputs are scored with a discriminative loss. The authors report that MTGR-large, trained on six months of data, increases conversion volume by 1.22% and CTR by 1.31% over a DLRM baseline that had been continuously optimized for two years, while reducing inference cost by 12%. They also observe power-law scaling of offline and online metrics with computational complexity across three model sizes.

Load-bearing premise

The paper's premise that removing cross features cannot be compensated by scaling is supported only by an ablation on the smallest model, so the claim that cross features remain indispensable at larger scales is assumed rather than directly measured.

Editorial extensions

If this is right

  • MTGR's offline and online metrics improve smoothly as the number of HSTU blocks, model width, and sequence length grow, consistent with a power-law scaling law in compute.
  • Retaining cross features in candidate tokens yields a significant gain over removing them; the paper reports that removing cross features erases MTGR-large's advantage over the DLRM baseline.
  • User-level aggregation makes inference cost sublinear in candidate count, allowing the deployed model to cut inference cost by 12% while using 65x more FLOPs per sample than the DLRM baseline.
  • Training MTGR-large on six months of data at roughly the same training cost as the production DLRM beats a DLRM model that has been continuously learned for two years, both offline and online.

Reading between the lines

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

  • If the power-law relationship holds as compute grows, MTGR points toward a single large ranking model shared across multiple scenarios, a direction the authors mention for future work; one would expect cross features to remain necessary even at foundation-model scale.
  • Because public recommender datasets usually lack handcrafted cross features, a direct transfer of MTGR to them may not reproduce the reported gains; constructing synthetic cross features would provide a testable bridge to academic benchmarks.
  • The dynamic masking rules for mixing static, real-time, and candidate tokens could transfer to other ranking or forecasting problems where past and future information must be separated without leakage.
  • The reported 65x FLOPs is per sample, and since one MTGR sample now contains many candidates, a per-candidate FLOPs comparison would give a clearer efficiency picture relative to DLRM.
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 / 6 minor

Summary. The paper proposes MTGR, an industrial ranking model that combines the HSTU-style transformer encoder of generative recommendation models with the handcrafted features of a traditional DLRM. The key design choices are: (i) converting user, sequence, real-time, and candidate features into tokens and aggregating all candidates of a user into one forward pass; (ii) preserving cross features by attaching them to candidate tokens; (iii) two proposed components, Group-Layer Normalization and a dynamic masking scheme to prevent temporal leakage; and (iv) training-system optimizations on TorchRec. The authors report offline and online gains over a production DLRM baseline (UserTower-SIM), three model sizes, and claim a power-law scaling relationship with compute, with MTGR-large deployed at Meituan.

Significance. MTGR is a credible industrial architecture that addresses a real limitation of pure generative recommenders (loss of handcrafted cross features) and the limitations of DLRM scaling. The paper has concrete strengths: a deployed system with A/B tests on production traffic, a consistent offline/online trend on CTCVR GAUC and UV_CTCVR across three sizes, ablations showing the value of GLN and dynamic masking, and detailed engineering contributions (dynamic hash tables, load balancing, embedding deduplication, mixed precision and fused attention kernels). If the power-law and cross-feature-indispensability claims were supported by the same measurement standard as the architecture description, this would be a valuable addition to the industrial recommendation literature. As it stands, the empirical evidence is not yet sufficient to establish those two strong scientific claims; the strengths are mostly architectural and systems-level.

major comments (4)
  1. [§6.4, Fig. 3(d)] The paper's headline scaling claim is presented as a 'power-law relationship between performance and computational complexity' (Section 6.4), but the evidence is at most three MTGR configurations (Table 2) plotted without a fitted equation, exponent, confidence intervals, or repeated runs. With three points, a monotone trend cannot distinguish a power law from any smooth curve, and the online PV_CTR column in Table 5 is non-monotonic (medium +2.29%, large +1.90%), which further weakens the universality of the claim. Please report the fitted form (e.g., gain = a·C^b) with its exponent and goodness of fit, add intermediate configurations or repeated seeds if possible, or explicitly restrict the claim to a qualitative monotone trend.
  2. [§6.3, Table 4] The abstract and Section 1 assert that removing cross features 'cannot be compensated by scaling up at all,' but the only ablation without cross features is run on MTGR-small. The sentence in Section 6.3 that this drop 'even erases the gain of MTGR-large over DLRM' is an extrapolation: MTGR-large was never trained without cross features. The extrapolation is numerically plausible (the small-model drop of 0.0137 in CTR GAUC exceeds MTGR-large's gain of 0.0073 over UserTower-SIM), but it is not a measured result. Please run the ablation at least at MTGR-medium/large, or revise the abstract and Section 1 to say that scaling did not compensate at the tested scale.
  3. [§6.5, Table 5] The text states that the online metrics demonstrate scalability and that the benefits 'continue to amplify' with training tokens, but the PV_CTR column in Table 5 is non-monotonic: MTGR-medium improves +2.29% while MTGR-large improves +1.90%. Even if UV_CTCVR is the primary business metric, the paper should explain this discrepancy or remove PV_CTR from the scalability claim.
  4. [§5 and Table 2] The claim of '65x FLOPs for single-sample forward inference' is ambiguous about the unit of comparison. In the DLRM baseline, an example is one (user, candidate) pair, while an MTGR example aggregates all candidates in a request (Eq. 3). 'GFLOPs/example' in Table 2 therefore does not directly compare total inference cost per request unless the number of candidates is specified. Please define the denominator and, if possible, report total FLOPs per request with the candidate count, so that the inference-cost reduction stated in Section 6.5 can be verified.
minor comments (6)
  1. [§6.3] The text says the ablation results are shown in 'Fig 4', but the results are in Table 4; the figure reference should be corrected.
  2. [§4.2, Eq. (5)] The scaling factor in the attention update is described as 'divided by total length of input features,' but the notation does not show a square-root or normalization convention; please state the exact denominator and whether this is a deliberate average-pooling choice.
  3. [§6.1] The learning rate is decreased as model size grows and the embedding dimensions are enlarged 'proportionally,' but no sensitivity analysis or selection criterion is given; because these choices affect the GFLOPs-versus-quality curve, a sentence explaining how these values were chosen would strengthen the scaling interpretation.
  4. [§3.1] There are several typos and notation issues: 'closet interaction' should be 'closest interaction'; the maximum length of \vec{S} is 1000 while \vec{R} is 100 (missing 'length'); and Figure 3's caption contains 'numer' and garbled math.
  5. [Abstract/§6.5] The abstract's claim of 'the largest gain in nearly two years both offline and online' is not supported by any time-series comparison in the paper. Please either provide the historical trend or remove the phrase.
  6. [§4.2] The paper says 'Features of the same domain (for example, U) form a group' for Group-Layer Normalization, but it does not specify the exact grouping for sequence tokens versus real-time tokens, or whether the same grouping is used at every layer; please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: MTGR's predictions are measured against external DLRM baselines, and the self-citation concerns non-load-bearing infrastructure.

full rationale

The paper's central claims are empirically evaluated rather than derived from their own assumptions. MTGR's HSTU backbone is cited from external work [23]; the ranking architecture, Group-Layer Normalization, and dynamic masking are evaluated by offline and online experiments against DLRM baselines. The reported power-law scaling relationship is an empirical fit to measured performance, not a definition of that performance, so it does not make the predictions equivalent to its inputs. The only self-citation, [21] MTGRBoost, supports the decoupled hash table engineering design, which is infrastructure optimization and not a load-bearing scientific premise of the recommendation model. The text in Section 6.3 extrapolates the MTGR-small 'w/o cross features' ablation to claim the drop would erase MTGR-large's gain over DLRM; this is an unmeasured extrapolation and a correctness/evidence concern, but it is not circularity because the extrapolation does not define or force the measured gains. No equation or fitted parameter is renamed as a prediction, and no central claim reduces by construction to an earlier self-derived result.

Assumptions & free parameters 1 free parameters · 3 assumptions · 2 invented entities

The central claim rests mainly on the availability of proprietary Meituan data and the assumption that cross features matter at scale. The paper's new components (GLN, dynamic masking) are ablated, but only on the smallest model and on internal data, so they lack independent evidence. The claimed power-law scaling is fit to three configurations and is not a parameter-free derivation.

free parameters (1)
  • power-law scaling fit = not reported
    Figure 3(d) claims a power-law relationship between CTCVR GAUC gain and FLOPs, fitted to three MTGR configurations; the equation, exponent, and error bars are not given.
assumptions (3)
  • domain assumption Cross features are crucial for ranking performance in Meituan's scenario
    Asserted in Sections 1 and 6.1; the only evidence is an ablation on MTGR-small and the authors' industrial experience, not a public benchmark.
  • domain assumption HSTU-style transformer architecture is an effective base for modeling user behavior sequences
    Adopted from cited work [23], not re-derived or compared against other generative architectures in this paper.
  • domain assumption User-level sample aggregation does not introduce harmful training bias
    The training system aggregates all candidates for a user into one sample (Section 4.1); the paper does not analyze distribution shift versus pointwise training.
invented entities (2)
  • Group-Layer Normalization (GLN)
    purpose: Normalize tokens from different semantic domains separately in the HSTU encoder to improve multi-domain encoding
    Introduced in this paper; ablation in Table 4 shows a small gain (CTR AUC 0.7631 vs 0.7606 without GLN), but no external validation.
  • Dynamic masking strategy
    purpose: Prevent information leakage between candidate tokens and real-time user actions while allowing full attention on static features
    Introduced in this paper; ablation in Table 4 shows a small gain (CTR AUC 0.7631 vs 0.7620 without dynamic mask), but no external validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MTGR: Industrial-Scale Generative Recommendation Framework in Meituan." pith.science (2026). https://pith.science/paper/LZX5FAUX

@misc{pith2026250518654,
  author       = {Pith},
  title        = {Pith review of: MTGR: Industrial-Scale Generative Recommendation Framework in Meituan},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LZX5FAUX}},
  note         = {Machine review of arXiv:2505.18654}
}
read the original abstract

Scaling law has been extensively validated in many domains such as natural language processing and computer vision. In the recommendation system, recent work has adopted generative recommendations to achieve scalability, but their generative approaches require abandoning the carefully constructed cross features of traditional recommendation models. We found that this approach significantly degrades model performance, and scaling up cannot compensate for it at all. In this paper, we propose MTGR (Meituan Generative Recommendation) to address this issue. MTGR is modeling based on the HSTU architecture and can retain the original deep learning recommendation model (DLRM) features, including cross features. Additionally, MTGR achieves training and inference acceleration through user-level compression to ensure efficient scaling. We also propose Group-Layer Normalization (GLN) to enhance the performance of encoding within different semantic spaces and the dynamic masking strategy to avoid information leakage. We further optimize the training frameworks, enabling support for our models with 10 to 100 times computational complexity compared to the DLRM, without significant cost increases. MTGR achieved 65x FLOPs for single-sample forward inference compared to the DLRM model, resulting in the largest gain in nearly two years both offline and online. This breakthrough was successfully deployed on Meituan, the world's largest food delivery platform, where it has been handling the main traffic.

Figures

Figures reproduced from arXiv: 2505.18654 by the authors.

Figure 1
Figure 1. Data Arrangement and workflow of traditional ranking [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Data Arrangement and architecture of MTGR. (a) represents data arrangement and overall workflow of MTGR: 3 candidates’ fea￾tures are aggregated with the counterpart of one user. The features are embedded and converted to token by MLP, forming a series of input sequence for self-attention with mask. The representation of the tokens of candidates are used for logit via another MLP module. (b) Detailed description of s… view at source ↗
Figure 3
Figure 3. MTGR performance improves smoothly as we increase the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 6 Pith papers

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

  1. Request-Only Optimization for Recommendation Systems

    cs.IR 2025-07 conditional novelty 6.0 of 10

    A request-level training data format eliminates duplicate user features, increasing storage efficiency and training throughput while enabling larger recommendation architectures.

  2. CCFormer: Efficient Cross-Field Interaction and Hierarchical Sequence Compression for Industrial Recommendation at Tencent

    cs.IR 2026-07 conditional novelty 5.0 of 10

    CCFormer beats strong industrial sequential rankers offline and online by pairing field-separated cross-attention with subspace token mixing and hierarchical sequence compression.

  3. An Industrial-Scale Sequential Recommender for LinkedIn Feed Ranking

    cs.IR 2026-02 conditional novelty 4.0 of 10

    A transformer-based sequential recommender, Feed SR, improved LinkedIn Feed time spent by 2.10% in an online A/B test and now serves the majority of Feed traffic.

  4. Coarse-to-Fine Long-term Interest Modeling for Generative Recommendation

    cs.IR 2026-02 conditional novelty 4.0 of 10

    GLASS extends generative retrieval with a tiered long-term interest vector and a first-SID-keyed search of long histories, reporting consistent gains over Tiger and DualGR on two public datasets.

  5. A Survey of Real-World Recommender Systems: Challenges, Constraints, and Industrial Perspectives

    cs.IR 2025-09 conditional novelty 4.0 of 10

    A survey of A/B-validated industrial recommender systems, split into transaction-oriented and content-oriented categories, with a discussion of the academia-industry gap.

  6. GR-LLMs: Recent Advances in Generative Recommendation Based on Large Language Models

    cs.IR 2025-07 unverdicted novelty 3.0 of 10

    A survey of LLM-based generative recommendation systems, covering application settings, training pipelines, industrial deployment challenges, and future directions.

Reference graph

Works this paper leans on

28 extracted references · 5 canonical work pages · cited by 6 Pith papers

  1. [21]

    Yuxiang Wang, Xiao Yan, Chi Ma, Mincong Huang, Xiaoguang Li, Lei Yu, Chuan Liu, Ruidong Han, He Jiang, Bin Yin, et al. 2025. MTGRBoost: Boosting Large-scale Generative Recommendation Models in Meituan. arXiv preprint arXiv:2505.12663 (2025)

    arXiv 2025

  2. [1]

    Qiwei Chen, Changhua Pei, Shanshan Lv, Chao Li, Junfeng Ge, and Wenwu Ou

  3. [2]

    Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. 2022. Flashat- tention: Fast and memory-efficient exact attention with io-awareness. Advances in neural information processing systems 35 (2022), 16344–16359

    2022

  4. [3]

    Jiaxin Deng, Shiyao Wang, Kuo Cai, Lejian Ren, Qigen Hu, Weifeng Ding, Qiang Luo, and Guorui Zhou. 2025. OneRec: Unifying Retrieve and Rank with Generative Recommender and Iterative Preference Alignment. arXiv preprint arXiv:2502.18965 (2025)

    arXiv 2025

  5. [4]

    Yan Fang, Jingtao Zhan, Qingyao Ai, Jiaxin Mao, Weihang Su, Jia Chen, and Yiqun Liu. 2024. Scaling laws for dense retrieval. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval . 1339–1349

    2024

  6. [5]

    Cesar Ferri, José Hernández-Orallo, and Peter A Flach. 2011. A coherent in- terpretation of AUC as a measure of aggregated classification performance. In Proceedings of the 28th International Conference on Machine Learning (ICML-11) . 657–664

    work page 2011

  7. [6]

    Xingzhuo Guo, Junwei Pan, Ximei Wang, Baixu Chen, Jie Jiang, and Mingsheng Long. 2023. On the embedding collapse when scaling up recommendation models. arXiv preprint arXiv:2310.04400 (2023)

    arXiv 2023

  8. [7]

    Ruidong Han, Qianzhong Li, He Jiang, Rui Li, Yurou Zhao, Xiang Li, and Wei Lin. 2024. Enhancing CTR Prediction through Sequential Recommendation Pre- training: Introducing the SRP4CTR Framework. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management . 3777–3781

    2024

Show all 28 references
  1. [8]

    Dmytro Ivchenko, Dennis Van Der Staay, Colin Taylor, Xing Liu, Will Feng, Rahul Kindi, Anirudh Sudarshan, and Shahin Sefati. 2022. Torchrec: a pytorch domain library for recommendation systems. In Proceedings of the 16th ACM Conference on Recommender Systems. 482–483

    2022

  2. [9]

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. 2020. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361 (2020)

    2020 arXiv

  3. [10]

    Mario Michael Krell, Matej Kosec, Sergio P Perez, and Andrew Fitzgibbon. 2021. Efficient sequence packing without cross-contamination: Accelerating large lan- guage models without impacting performance. arXiv preprint arXiv:2107.02027 (2021)

    2021 arXiv

  4. [11]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning . PMLR, 19730–19742

    2023

  5. [12]

    Jianxun Lian, Xiaohuan Zhou, Fuzheng Zhang, Zhongxia Chen, Xing Xie, and Guangzhong Sun. 2018. xdeepfm: Combining explicit and implicit feature in- teractions for recommender systems. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data ...

    2018

  6. [13]

    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. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining . 1930–1939

    2018

  7. [14]

    William Peebles and Saining Xie. 2023. Scalable diffusion models with transform- ers. In Proceedings of the IEEE/CVF international conference on computer vision . 4195–4205

    2023

  8. [15]

    Qi Pi, Guorui Zhou, Yujing Zhang, Zhe Wang, Lejian Ren, Ying Fan, Xiaoqiang Zhu, and Kun Gai. 2020. Search-based user interest modeling with lifelong sequential behavior data for click-through rate prediction. In Proceedings of the 29th ACM International Conference on Informat...

    2020

  9. [16]

    Kyuyong Shin, Hanock Kwak, Su Young Kim, Max Nihlén Ramström, Jisu Jeong, Jung-Woo Ha, and Kyung-Min Kim. 2023. Scaling law for recommendation models: Towards general-purpose user representations. In Proceedings of the AAAI conference on artificial intelligence , Vol. 37. 4596–4604

    2023

  10. [17]

    Zihua Si, Lin Guan, ZhongXiang Sun, Xiaoxue Zang, Jing Lu, Yiqun Hui, Xingchao Cao, Zeyu Yang, Yichen Zheng, Dewei Leng, et al. 2024. Twin v2: Scaling ultra- long user behavior sequence modeling for enhanced ctr prediction at kuaishou. In Proceedings of the 33rd ACM Internatio...

    2024

  11. [18]

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

    2020

  12. [19]

    Ruoxi Wang, Rakesh Shivanna, Derek Cheng, Sagar Jain, Dong Lin, Lichan Hong, and Ed Chi. 2021. Dcn v2: Improved deep & cross network and practical lessons for web-scale learning to rank systems. In Proceedings of the web conference 2021 . 1785–1797

    2021

  13. [20]

    Xu Wang, Jiangxia Cao, Zhiyi Fu, Kun Gai, and Guorui Zhou. 2024. HoME: Hierarchy of Multi-Gate Experts for Multi-Task Learning at Kuaishou. arXiv preprint arXiv:2408.05430 (2024)

    2024 arXiv

  14. [22]

    Bencheng Yan, Shilei Liu, Zhiyuan Zeng, Zihao Wang, Yizhen Zhang, Yujin Yuan, Langming Liu, Jiaqi Liu, Di Wang, Wenbo Su, et al. 2025. Unlocking Scaling Law in Industrial Recommendation Systems with a Three-step Paradigm based Large User Model. arXiv preprint arXiv:2502.08309 (2025)

    2025

  15. [23]

    Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhao- jie Gong, Fangda Gu, Michael He, et al. 2024. Actions speak louder than words: Trillion-parameter sequential transducers for generative recommendations.arXiv preprint arXiv:2402.17152 (2024)

    2024 arXiv

  16. [24]

    Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. 2022. Scaling vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 12104–12113

    2022

  17. [25]

    Buyun Zhang, Liang Luo, Yuxin Chen, Jade Nie, Xi Liu, Daifeng Guo, Yanli Zhao, Shen Li, Yuchen Hao, Yantao Yao, et al. 2024. Wukong: Towards a scaling law for large-scale recommendation. arXiv preprint arXiv:2403.02545 (2024)

    2024 arXiv

  18. [26]

    Wei Zhang, Dai Li, Chen Liang, Fang Zhou, Zhongke Zhang, Xuewei Wang, Ru Li, Yi Zhou, Yaning Huang, Dong Liang, et al. 2024. Scaling User Modeling: Large-scale Online User Representations for Ads Personalization in Meta. In Companion Proceedings of the ACM Web Conference 2024 . 47–55

    2024

  19. [27]

    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. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining ...

    2018

  20. [2021]

    arXiv preprint arXiv:2108.04468 (2021)

    End-to-end user behavior retrieval in click-through rateprediction model. arXiv preprint arXiv:2108.04468 (2021)

    2021 arXiv

Pith tools

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