Pith. sign in

REVIEW 5 major objections 5 minor 75 references

Large Foundation Model for Ads Recommendation

T0 review · 5 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read LFM4Ads claims that moving user, item, and cross representations from a pre-trained foundation model into downstream ad models lifts platform GMV by 2.45%.

desk verdict A substantial industrial LFM-for-ads transfer system whose architecture is worth studying, but whose headline 2.45% GMV lift is a point estimate from production A/B tests that cannot be cleanly attributed to the framework. read the letter →

arxiv 2508.14948 v1 pith:DGGOOQJY submitted 2025-08-20 cs.LG

Shangyu Zhang , Shijie Quan , Zhongren Wang , Junwei Pan , Tianqu Zhuang , Bo Fu , Yilong Sun , Jieying Lin
show 25 more authors
Jushuo Chen Xiaotian Li Zhixiang Feng Xian Hu Huiting Deng Hua Lu Jinpeng Wang Boqi Dai Xiaoyu Chen Bin Hu Lili Huang Yanwen Wu Yeshou Cai Qi Zhou Huang Tang Chunfeng Yang Chengguo Yin Tingyu Jiang Lifeng Wang Shudong Huang Dapeng Liu Lei Xiao Haijie Gu Shu-Tao Xia Jie Jiang
This is my paper · ORCID
classification cs.LG
keywords LFM4Adsadsrecommendationfoundationmodeltransferlearningcrossrepresentationmulti-granularityclick-throughratepredictionindustrialdeployment
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that a large pre-trained foundation model can improve online advertising recommendation not by transferring only user interest vectors, but by transferring all three kinds of representations it learns — user, item, and user-item cross representations — and by using each at the right granularity in downstream models: as a feature, as a reusable module, or as a standalone retrieval model. If true, this means the industry's upstream-downstream paradigm for ads has been leaving value on the table by ignoring item and cross embeddings. The paper backs the claim with production experience at Tencent: over ten launches since Q4 2024, an overall 2.45% GMV lift across the platform. The central question it wants settled is whether moving all representations, not just user vectors, is what makes large foundation models pay off in ads.

What carries the argument

The triple-tower design with shared user/item towers and a dual-branch mix tower (content branch and ad branch); the time-interval decaying EMA aggregation that converts sample-level CRs into coarse-grained user-CR and item-CR vectors; and the three usage mechanisms — non-linear interaction (gating with a trainable projection and sigmoid), the isomorphic interaction module (architectural reuse of the mix tower's interaction-and-DNN structure), and standalone retrieval (cosine similarity with a tiny adapter). The dual-branch mix tower is what separates commercial from non-commercial interest and is reported as critical in ablation.

What would settle it

Rerun the Table 1 baseline-vs-LFM4Ads contrast on identical live traffic with a fixed metric plan, per-bucket confidence intervals, and a holdout that freezes both models mid-experiment; if the 0.0005 AUC difference between SUM and LFM4Ads falls within the interval, the all-representation claim is unsupported. Meanwhile, inspect the per-launch breakdown: if the 2.45% aggregate GMV lift is carried mostly by one pCVR scenario while pCTR lifts are flat, the broad 'ads recommendation' benefit is weakened.

Watch

Extended reading notes

Core claim

LFM4Ads is a triple-tower architecture whose user and item towers produce UR and IR, and whose mix tower produces a user-item cross representation CR. The paper's central discovery is that CRs, unlike URs, are not directly transferable: they live at the sample level, tied to a particular user-item pair, so they must be aggregated over time into user-level and item-level forms via a time-interval decaying exponential moving average. It also finds empirically that the penultimate MLP layer of the mix tower is the most transferable place to extract CRs. Downstream, the representations are deployed in three modes: a non-linear gating adapter that lets a CR modulate downstream features, an isomor

Load-bearing premise

The headline numbers depend on the online A/B tests at Tencent being unconfounded: if production changes, traffic shifts, or metric choices other than LFM4Ads contributed to the reported AUC and GMV lifts, the causal claim collapses.

Editorial extensions

If this is right

  • Existing user-representation-only LFM pipelines can be upgraded by adding item and cross-representation transfer, likely with gains concentrated in cold-start scenarios.
  • Downstream ranking models can be thinned: module-level and model-level transfer reduce the parameters that need to be re-learned per task.
  • Cross-domain pretraining (content plus ads) is shown to improve ad ranking even in a zero-baseline comparison, supporting unified pretraining beyond a single domain.
  • The finding that intermediate MLP layers are the most transferable extraction point for cross representations gives a practical recipe for other large recommendation models.
  • The reported 0.005 AUC gain over SUM, if measured cleanly, exceeds the 0.001-AUC industrial significance threshold cited from prior work.

Reading between the lines

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

  • The time-decaying EMA aggregation of pair-level signals into entity-level states may transfer to other online-learning systems that must summarize frequent, non-stationary interactions; this is an editorial extension.
  • The 'stronger the foundation, the thinner the downstream models' thesis, stated in the paper's introduction, implies a roadmap where downstream tasks eventually shrink to heads and retrieval scorers — a trajectory the three usage modes sketch but do not fully test.
  • The standalone retrieval deployment's reported engagement gains (CTR up, quick-slip down) suggest model-level transfer may improve user experience beyond GMV; a testable extension is mixing retrieval results into the ranking stage rather than keeping them separate.
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

5 major / 5 minor

Summary. The paper proposes LFM4Ads, a large foundation model for ads recommendation deployed at Tencent. It uses a triple-tower architecture to extract user, item, and user-item cross representations, aggregates cross representations via a time-interval-decaying exponential moving average, and transfers them into downstream tasks in three granularities: feature-level non-linear interaction, module-level isomorphic interaction, and model-level standalone retrieval. The paper reports online A/B evaluations, including a platform-wide GMV lift of 2.45% from 10+ production launches, with the main comparison against SUM and several ablations.

Significance. If the empirical claims hold, this is a significant industrial contribution. The all-representation transfer design and the feature/module/model decomposition are clearly described and fill a real gap relative to user-representation-only pre-training. The deployment scale (terabyte-scale parameters, tens of billions of daily samples) and the inclusion of negative internal results for linear/projection fusion are valuable. However, the central claims are self-reported online measurements with no statistical details; the scientific content is a design plus a plausibility argument rather than a verified causal evaluation. The paper would be strengthened considerably by releasing de-aggregated, per-launch experimental evidence.

major comments (5)
  1. [Section 5.1.1, Table 1; Section 5.2] The headline results are point estimates without confidence intervals, significance tests, experiment durations, or bucket sizes. Table 1 reports AUC to four decimals and GMV lifts without variance; Section 5.2 reports an aggregate 2.45% GMV lift across 10+ launches. The paper's own Sec 5.1.2 says the LFM went through three iterative versions, and Sec 4.3.1 describes continuous online updates, so concurrent changes are plausible confounds. The assertion that a 0.001 AUC lift is significant is justified by a self-citation [34] with no power analysis. Please provide per-launch statistical details and a decomposition of the aggregate, or explicitly state that the aggregate is not from a controlled experiment.
  2. [Section 5.1.3, Fig. 5] The optimal CR extraction layer is selected using the same online metric family as the reported final win. Fig. 5 shows AUC lifts without error bars or repeated trials, and the conclusion that intermediate DNN layers are most transferable is based on linear fusion only, whereas the deployed feature-level usage is the non-linear interaction of Eq. (3). A layer choice that is optimal for linear probes need not be optimal for the deployed interaction, and selection on the evaluation metric risks overfitting. Report the selection protocol and, ideally, validate on held-out scenarios.
  3. [Section 4.3.3] The Light-weight Embedding Evaluator (LEE) is used to select among LFM variants by replacing raw cross representations with proxy models and monitoring AUC. No evidence is given that LEE's proxy results are calibrated with, or even rank-correlated with, full-scale online A/B outcomes. Because LEE drives the iterative version choices that produce the reported gains, its validity is load-bearing. Provide a validation study comparing LEE decisions with full online experiments on a set of past variants.
  4. [Section 5.2] The per-launch GMV numbers are not reconciled with the platform-wide 2.45% figure. Feature-level launches report lifts of 0.42%, 2.53%, 0.70%, 1.75%, 0.76%, and 0.93%; module-level usage reports 1.88%; standalone retrieval is GMV-neutral. The paper does not state traffic shares, overlap between launches, time periods, or whether the 2.45% is a sum, a weighted average, or a net after exclusions. A transparent decomposition is necessary for the headline number to be interpretable.
  5. [Section 5.1.1; Introduction] The abstract and introduction claim 'consistent improvements over state-of-the-art methods', but the only external comparison in Table 1 is SUM (plus ablations of the proposed components). No comparison is made to other recent LFM-based recommendation systems such as PinFM [6], PinnerFormer [36], or ItemSage [1]. The claimed SOTA status is therefore unsupported by the reported experiments. Please either add comparisons or soften the claim to 'improvement over SUM'.
minor comments (5)
  1. [Section 4.3.2 and intro] Typo: 'false torerence' in the intro and Section 4.3.2 should be 'false tolerance'.
  2. [Conclusion and Section 5.3.1] The name is inconsistent: the conclusion uses 'LLM4Ads' and Section 5.3.1 uses 'LFM4Rec'; should be 'LFM4Ads' throughout.
  3. [Section 3.4.1, Eq. (3)] The dimensions in Eq. (3) are ambiguous: the text says d and d′ are the dimensions of CR(u) and e_i, but the Hadamard product requires sigma(CR(u)M) to have the same dimension as e_i. Please clarify the shapes.
  4. [Table 1] The rows marked 'No Launch' should be explained. Was the configuration ever deployed in a live experiment? What criterion was used to decide not to launch (e.g., metric threshold, significance)? Without this, the reader cannot tell whether 'No Launch' means 'failed' or 'not attempted'.
  5. [Figure 5] The figure lacks axis labels for the 'various output dimensions of 16, 32, 64' and does not indicate which scenario each panel corresponds to. Please label the panels and axes.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: central claims are online measurements, not derived predictions; one minor self-citation for an AUC-significance heuristic is not load-bearing.

full rationale

This is an industrial systems paper without a parametric derivation chain: the reported lifts (Table 1 and Sec. 5.2) are measured online A/B outcomes, not quantities obtained by fitting a parameter to a subset of data and then predicting a closely related quantity. The design choices (dual branch, extraction layer, non-linear fusion, aggregation) were selected using the same downstream metrics that are later reported, which creates a mild selection-on-test-metric burden, but the paper does not present these selections as out-of-sample predictions or as a first-principles derivation. The only notable self-referential element is in Sec. 5.1.1, where the convention 'a 0.001 AUC lift is consider significant in industrial recommenders [34]' is attributed to the authors' own prior KDD paper. This citation supports the interpretation that the 0.005 AUC gain is meaningful, but the central efficacy claims also rest on independent online GMV measurements (1.35% in Table 1, 2.45% platform-wide in Sec. 5.2). Accordingly, the self-citation is minor and not load-bearing for the headline result. No equation is equivalent by construction to its inputs; no fitted parameter is renamed as a prediction. The main vulnerabilities are evaluation confounds (unreported error bars, durations, bucket sizes, and concurrent production changes), which are correctness/evidence concerns rather than circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

This is an industrial systems paper; the central claim depends on empirical assumptions (A/B test validity, data alignment) and unspecified hyperparameters (decay function, layer index, projection dimensions) rather than on a formal derivation. The ledger records these dependencies.

free parameters (4)
  • beta(t), time-interval decay function in CR aggregation = not specified; described as a non-increasing function
    Equations (1) and (2) depend on beta(t), but its functional form and parameters are not given. The reported performance therefore depends on an unstated choice.
  • CR extraction layer index = penultimate DNN layer (selected by online experiments)
    Section 3.3 and 5.1.3 say the penultimate MLP layer performs best; this layer index is chosen using the same downstream AUC metrics used for the headline claim.
  • projection matrix M in non-linear interaction (Eq. 3) = trainable; dimensions d' unspecified
    The matrix is learned, and its output dimension d' is not specified, making exact implementation and replication impossible.
  • adapters in standalone retrieval (Eq. 6) = trainable MLPs; architecture unspecified
    The adapter networks are described only as tiny MLPs; their structure and dimensions are not given.
assumptions (4)
  • domain assumption Online A/B test metrics are measured correctly and the reported lifts are attributable to the model change.
    The entire central claim (2.45% GMV lift) rests on self-reported production A/B tests in Sections 5.1 and 5.2, with no independent validation or statistical details.
  • domain assumption Content and ad domains share alignable user interests via a unified commercial tag schema.
    Section 4.1 uses LLM-generated synthetic tags and a 7B-parameter multi-modal LLM to align content and ad features; if this alignment fails, the cross-domain pre-training benefit is not established.
  • domain assumption The dual-branch mix tower decouples commercial from non-commercial interests and avoids negative transfer.
    Section 5.1.1 compares same-branch and dual-branch; the superiority of dual-branch is based on a single ablation series without significance testing.
  • standard math Standard large-scale deep learning assumptions hold: optimization converges, embeddings are stable, and feature engineering is adequate.
    The paper does not prove any convergence or stability properties; it assumes typical CTR model training behavior, and reports volatility measurements in Section 5.3.2 that support stability empirically but only for one 5-minute window.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Large Foundation Model for Ads Recommendation." pith.science (2026). https://pith.science/paper/DGGOOQJY

@misc{pith2026250814948,
  author       = {Pith},
  title        = {Pith review of: Large Foundation Model for Ads Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DGGOOQJY}},
  note         = {Machine review of arXiv:2508.14948}
}
read the original abstract

Online advertising relies on accurate recommendation models, with recent advances using pre-trained large-scale foundation models (LFMs) to capture users' general interests across multiple scenarios and tasks. However, existing methods have critical limitations: they extract and transfer only user representations (URs), ignoring valuable item representations (IRs) and user-item cross representations (CRs); and they simply use a UR as a feature in downstream applications, which fails to bridge upstream-downstream gaps and overlooks more transfer granularities. In this paper, we propose LFM4Ads, an All-Representation Multi-Granularity transfer framework for ads recommendation. It first comprehensively transfers URs, IRs, and CRs, i.e., all available representations in the pre-trained foundation model. To effectively utilize the CRs, it identifies the optimal extraction layer and aggregates them into transferable coarse-grained forms. Furthermore, we enhance the transferability via multi-granularity mechanisms: non-linear adapters for feature-level transfer, an Isomorphic Interaction Module for module-level transfer, and Standalone Retrieval for model-level transfer. LFM4Ads has been successfully deployed in Tencent's industrial-scale advertising platform, processing tens of billions of daily samples while maintaining terabyte-scale model parameters with billions of sparse embedding keys across approximately two thousand features. Since its production deployment in Q4 2024, LFM4Ads has achieved 10+ successful production launches across various advertising scenarios, including primary ones like Weixin Moments and Channels. These launches achieve an overall GMV lift of 2.45% across the entire platform, translating to estimated annual revenue increases in the hundreds of millions of dollars.

Figures

Figures reproduced from arXiv: 2508.14948 by the authors.

Figure 1
Figure 1. The comparison between existing works and our [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The architecture of LFM4Ads, consisting of a con [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Three usages of our representations: (left) feature [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Our online system. Whenever an advertisement [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: AUC lift in two downstream scenarios with various [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: t-SNE and K-means Visualization of Proposed Em [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Embedding Kmeans-based clustering. the top three behaviors and their proportions in the visualization. We found: (1) Clusters 0, 1, 6, and 8 were grouped together, repre￾senting users primarily engaged in activities related to childcare; (2) Cluster 4 aggregated users …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

75 extracted references · 46 canonical work pages

  1. [34]

    Junwei Pan, Wei Xue, Ximei Wang, Haibin Yu, Xun Liu, Shijie Quan, Xueming Qiu, Dapeng Liu, Lei Xiao, and Jie Jiang. 2024. Ads recommendation in a collapsed and entangled world. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 5566–5577

    2024

  2. [6]

    Xiangyi Chen, Kousik Rajesh, Matthew Lawhon, Zelun Wang, Hanyu Li, Haomiao Li, Saurabh Vishwas Joshi, Pong Eksombatchai, Jaewon Yang, Yi-Ping Hsu, et al

    work page

  3. [36]

    Nikil Pancha, Andrew Zhai, Jure Leskovec, and Charles Rosenberg. 2022. Pinner- Former: Sequence Modeling for User Representation at Pinterest. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’22). 3702–3712. https://doi.org/10.1145/3534678.3539156

    arXiv 2022

  4. [2]

    2024.IAB/PwC Internet Advertising Revenue Report

    Interactive Advertising Bureau. 2024.IAB/PwC Internet Advertising Revenue Report

    work page 2024

  5. [3]

    Zheng Chai, Qin Ren, Xijun Xiao, Huizhi Yang, Bo Han, Sijun Zhang, Di Chen, Hui Lu, Wenlin Zhao, Lele Yu, et al. 2025. LONGER: Scaling Up Long Sequence Modeling in Industrial Recommenders. arXiv preprint arXiv:2505.04421 (2025)

    arXiv 2025

  6. [4]

    Jianxin Chang, Chenbin Zhang, Zhiyi Fu, Xiaoxue Zang, Lin Guan, Jing Lu, Yiqun Hui, Dewei Leng, Yanan Niu, Yang Song, et al. 2023. TWIN: TWo-stage interest network for lifelong user behavior modeling in CTR prediction at kuaishou. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3785–3794

    2023

  7. [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. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining . 3795–3804

    work page 2023

  8. [7]

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

Show all 75 references
  1. [8]

    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

  2. [9]

    Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen Anderson, Greg Corrado, Wei Chai, Mustafa Ispir, Rohan Large Foundation Model for Ads Recommendation Conference’17, July 2017, Washington, DC, USA Anil, Zakaria Haque, Lichan Hong, V...

    2017

  3. [10]

    Ningya Feng, Junwei Pan, Jialong Wu, Baixu Chen, Ximei Wang, Qian Li, Xian Hu, Jie Jiang, and Mingsheng Long. 2024. Long-Sequence Recommendation Models Need Decoupled Embeddings. arXiv preprint arXiv:2410.02604 (2024)

    2024 arXiv

  4. [11]

    Mihajlo Grbovic and Haibin Cheng. 2018. Real-Time Personalization Using Embeddings for Search Ranking at Airbnb. InProceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD ’18) . 311–

    2018

  5. [13]

    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)

    2023 arXiv

  6. [14]

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

    Wide & Deep Learning for Recommender Systems. In Proceedings 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

  7. [15]

    Xinran He, Junfeng Pan, Ou Jin, Tianbing Xu, Bo Liu, Tao Xu, Yanxin Shi, Antoine Atallah, Ralf Herbrich, Stuart Bowers, et al. 2014. Practical lessons from predicting clicks on ads at facebook. In International Workshop on Data Mining for Online Advertising (ADKDD). 1–9

    2014

  8. [16]

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk

  9. [17]

    Jie Hu, Li Shen, and Gang Sun. 2018. Squeeze-and-excitation networks. InProceed- ings of the IEEE conference on computer vision and pattern recognition . 7132–7141

    2018

  10. [18]

    Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: a factorization-machine based neural network for CTR prediction.arXiv preprint arXiv:1703.04247 (2017)

    2017 arXiv

  11. [19]

    Yuchin Juan, Yong Zhuang, Wei-Sheng Chin, and Chih-Jen Lin. 2016. Field- aware factorization machines for CTR prediction. In Proceedings of the 10th ACM Conference on Recommender Systems (RecSys) . 43–50

    2016

  12. [20]

    Xiangnan He and Tat-Seng Chua. 2017. Neural factorization machines for sparse predictive analytics. In Proceedings of the 40th International ACM SIGIR conference on Research and Development in Information Retrieval . 355–364

    2017

  13. [21]

    Yu Kang, Junwei Pan, Jipeng Jin, Shudong Huang, Xiaofeng Gao, and Lei Xiao

  14. [22]

    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. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management...

    2020

  15. [23]

    Xiangyang Li, Bo Chen, Lu Hou, and Ruiming Tang. 2023. CTRL: Connect Tabular and Language Model for CTR Prediction. arXiv preprint arXiv:2306.02841 (2023)

    2023 arXiv

  16. [24]

    Yangning Li, Shirong Ma, Xiaobin Wang, Shen Huang, Chengyue Jiang, Hai-Tao Zheng, Pengjun Xie, Fei Huang, and Yong Jiang. 2024. Ecomgpt: Instruction- tuning large language models with chain-of-task tasks for e-commerce. In Pro- ceedings of the AAAI Conference on Artificial Int...

    2024

  17. [25]

    Tongwen Huang, Zhiqi Zhang, and Junlin Zhang. 2019. FiBiNET: combining fea- ture importance and bilinear feature interaction for click-through rate prediction. In Proceedings of the 13th ACM Conference on Recommender Systems . 169–177

    2019

  18. [26]

    Mingfu Liang, Xi Liu, Rong Jin, Boyang Liu, Qiuling Suo, Qinghai Zhou, Song Zhou, Laming Chen, Hua Zheng, Zhiyuan Li, et al . 2025. External large foun- dation model: How to efficiently serve trillions of parameters for online ads recommendation. In Companion Proceedings of th...

    2025

  19. [27]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In 2018 IEEE International Conference on Data Mining (ICDM) . IEEE, 197–206

    2018

  20. [28]

    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 KDD. ACM, 1930–1939

    2018

  21. [29]

    ECML-PKDD (2025)

    Towards Unifying Feature Interaction Models for Click-Through Rate Prediction. ECML-PKDD (2025)

    2025

  22. [30]

    Kelong Mao, Jieming Zhu, Liangcai Su, Guohao Cai, Yuru Li, and Zhenhua Dong

  23. [31]

    H Brendan McMahan, Gary Holt, David Sculley, Michael Young, Dietmar Ebner, Julian Grady, Lan Nie, Todd Phillips, Eugene Davydov, Daniel Golovin, et al. 2013. Ad click prediction: a view from the trenches. In ACM SIGKDD International conference on Knowledge Discovery & Data Min...

    2013

  24. [32]

    Junwei Pan, Yizhi Mao, Alfonso Lobos Ruiz, Yu Sun, and Aaron Flores. 2019. Predicting different types of conversions with multi-task learning in online advertising. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining . 2689–2697

    2019

  25. [33]

    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 and Dat...

    2018

  26. [35]

    Zhutian Lin, Junwei Pan, Haibin Yu, Xi Xiao, Ximei Wang, Zhixiang Feng, Shifeng Wen, Shudong Huang, Lei Xiao, and Jie Jiang. 2024. Crocodile: Cross Experts Covariance for Disentangled Learning in Multi-Domain Recommendation. arXiv preprint arXiv:2405.12706 (2024)

    2024 arXiv

  27. [37]

    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. In SIGIR. 1137–1140

    2018

  28. [38]

    Qi Pi, Weijie Bian, Guorui Zhou, Xiaoqiang Zhu, and Kun Gai. 2019. Practice on Long Sequential User Behavior Modeling for Click-Through Rate Prediction. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD ’19) . 2671–2679. h...

    2019 doi

  29. [39]

    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 ACM International Conference on Information & Knowledge Manageme...

    2020

  30. [40]

    Yanru Qu, Bohui Fang, Weinan Zhang, Ruiming Tang, Minzhe Niu, Huifeng Guo, Yong Yu, and Xiuqiang He. 2018. Product-based neural networks for user response prediction over multi-field categorical data. ACM Transactions on Information Systems (TOIS) 37, 1 (2018), 1–35

    2018

  31. [41]

    Steffen Rendle. 2010. Factorization Machines. In Proceedings of the 2010 IEEE International Conference on Data Mining (ICDM ’10) . 995–1000. https://doi.org/ 10.1109/ICDM.2010.127

    2010 doi

  32. [42]

    Junwei Pan, Jian Xu, Alfonso Lobos Ruiz, Wenliang Zhao, Shengjun Pan, Yu Sun, and Quan Lu. 2018. Field-weighted factorization machines for click-through rate prediction in display advertising. In Proceedings of the 2018 World Wide Web Conference (WWW). 1349–1357

    2018

  33. [43]

    Xiang-Rong Sheng, Feifan Yang, Litong Gong, Biao Wang, Zhangming Chan, Yujing Zhang, Yueyao Cheng, Yong-Nan Zhu, Tiezheng Ge, Han Zhu, et al

  34. [44]

    Junwei Pan, Zhilin Zhang, Han Zhu, Jian Xu, Jie Jiang, and Bo Zheng. 2025. Computational Advertising: Recent Advances. In Companion Proceedings of the ACM on Web Conference 2025. 37–40

    2025

  35. [45]

    Kaize Shi, Xueyao Sun, Dingxian Wang, Yinlin Fu, Guandong Xu, and Qing Li

  36. [46]

    Bo Peng, Xinyi Ling, Ziru Chen, Huan Sun, and Xia Ning. 2024. ecellm: Gen- eralizing large language models for e-commerce from large-scale, high-quality instruction data. arXiv preprint arXiv:2402.08831 (2024)

    2024 arXiv

  37. [47]

    Liangcai Su, Junwei Pan, Ximei Wang, Xi Xiao, Shijie Quan, Xihua Chen, and Jie Jiang. 2024. STEM: Unleashing the Power of Embeddings for Multi-task Recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 38. 9002–9010

    2024

  38. [48]

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

  39. [49]

    Yang Sun, Junwei Pan, Alex Zhang, and Aaron Flores. 2021. Fm2: Field-matrixed factorization machines for recommender systems. In Proceedings of the Web Conference 2021. 2828–2837

    2021

  40. [50]

    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. In RecSys. ACM, 269–278

    2020

  41. [51]

    Yuji Roh, Qingyun Liu, Huan Gui, Zhe Yuan, Yujin Tang, Steven Euijong Whang, Liang Liu, Shuchao Bi, Lichan Hong, Ed H Chi, et al. 2024. Levi: generalizable fine- tuning via layer-wise ensemble of different views.arXiv preprint arXiv:2402.04644 (2024)

    2024 arXiv

  42. [52]

    Jiancheng Wang, Mingjia Yin, Junwei Pan, Ximei Wang, Hao Wang, and Enhong Chen. 2025. Enhancing CTR Prediction with De-correlated Expert Networks. arXiv preprint arXiv:2505.17925 (2025)

    2025

  43. [53]

    InProceedings of the 33rd ACM International Conference on Information and Knowledge Management

    Enhancing Taobao Display Advertising with Multimodal Representations: Challenges, Approaches and Insights. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management . 4858–4865

  44. [54]

    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

  45. [55]

    Mingjia Yin, Junwei Pan, Hao Wang, Ximei Wang, Shangyu Zhang, Jie Jiang, Defu Lian, and Enhong Chen. 2025. From Feature Interaction to Feature Generation: A Generative Paradigm of CTR Prediction Models. In Forty-second International Conference on Machine Learning

    2025

  46. [56]

    arXiv preprint arXiv:2308.04913 (2023)

    LLaMA-E: empowering E-commerce authoring with object-interleaved instruction following. arXiv preprint arXiv:2308.04913 (2023)

    2023 arXiv

  47. [57]

    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. arXiv preprint arXiv:2407.16357 (2024)

    2024 arXiv

  48. [58]

    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 on Web Conference 20...

    2024

  49. [59]

    Guorui Zhou, Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, and Kun Gai. 2019. Deep interest evolution network for click-through rate prediction. In AAAI Conference on Artificial Intelligence (AAAI) , Vol. 33. 5941–5948

    2019

  50. [60]

    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 ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD) . 1059–1068

    2018

  51. [61]

    Haolin Zhou, Junwei Pan, Xinyi Zhou, Xihua Chen, Jie Jiang, Xiaofeng Gao, and Guihai Chen. 2024. Temporal Interest Network for User Response Prediction. In Companion Proceedings of the ACM on Web Conference 2024 . 413–422

    2024

  52. [62]

    Jie Zhou, Xianshuai Cao, Wenhao Li, Kun Zhang, Chuan Luo, and Qian Yu

  53. [63]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In NeurIPS. Conference’17, July 2017, Washington, DC, USA Shangyu Zhang, Shijie Quan, Zhongren Wang et al

    2017

  54. [64]

    Jie Zhu, Zhifang Fan, Xiaoxie Zhu, Yuchen Jiang, Hangyu Wang, Xintian Han, Haoran Ding, Xinmin Wang, Wenlin Zhao, Zhen Gong, et al. 2025. RankMixer: Scaling Up Ranking Models in Industrial Recommenders. arXiv preprint arXiv:2507.15551 (2025)

    2025 arXiv

  55. [65]

    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 (WWW). 1785–1797

    2021

  56. [66]

    Dongbo Xi, Zhen Chen, Peng Yan, Yinger Zhang, Yongchun Zhu, Fuzhen Zhuang, and Yu Chen. 2021. Modeling the sequential dependence among audience multi- step conversions with multi-task learning in targeted display advertising. In Proceedings of the 27th ACM SIGKDD Conference on...

    2021

  57. [68]

    Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. 2014. How transferable are features in deep neural networks? Advances in neural information processing systems 27 (2014)

    2014

  58. [69]

    Buyun Zhang, Liang Luo, Xi Liu, Jay Li, Zeliang Chen, Weilin Zhang, Xiaohan Wei, Yuchen Hao, Michael Tsang, Wenjun Wang, et al. 2022. DHEN: A deep and hierarchical ensemble network for large-scale click-through rate prediction. arXiv preprint arXiv:2203.11014 (2022)

    2022 arXiv

  59. [75]

    arXiv preprint arXiv:2303.06095 (2023)

    HiNet: A Novel Multi-Scenario & Multi-Task Learning Approach with Hierarchical Information Extraction. arXiv preprint arXiv:2303.06095 (2023)

    2023 arXiv

  60. [76]

    Zheli Zhou, Chenxu Zhu, Jianghao Lin, Bo Chen, Ruiming Tang, Weinan Zhang, and Yong Yu. 2025. Generative Representational Learning of Foundation Models for Recommendation. arXiv preprint arXiv:2506.11999 (2025)

    2025 arXiv

  61. [320]

    https://doi.org/10.1145/3219819.3219885

  62. [2015]

    arXiv preprint arXiv:1511.06939 (2015)

    Session-based recommendations with recurrent neural networks. arXiv preprint arXiv:1511.06939 (2015)

    2015 arXiv

  63. [2016]

    In Proceedings of the 1st workshop on deep learning for recommender systems

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

  64. [2019]

    In ACM International Conference on Information and Knowledge Management (CIKM)

    BERT4Rec: Sequential recommendation with bidirectional encoder repre- sentations from transformer. In ACM International Conference on Information and Knowledge Management (CIKM). 1441–1450

  65. [2023]

    FinalMLP: An Enhanced Two-Stream MLP Model for CTR Prediction.arXiv preprint arXiv:2304.00902 (2023)

    2023 arXiv

  66. [2024]

    https://www.iab.com/insights/internet-advertising-revenue-report-2024/

    2024

  67. [2025]

    arXiv preprint arXiv:2507.12704 (2025)

    PinFM: Foundation Model for User Activity Sequences at a Billion-scale Visual Discovery Platform. arXiv preprint arXiv:2507.12704 (2025)

    2025 arXiv

Pith tools

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