Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

LiveForesighter: Generating Future Information for Live-Streaming Recommendations at Kuaishou

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

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

Pith's one-line read Live-streaming recommendations can be improved by generating, rather than merely measuring, what a stream will do a few minutes ahead, via forecasted behavior statistics and next-product categories fed into the ranker.

desk verdict Plausible industrial application of forecasting to live-stream ranking, but the timing of the forecast window relative to labels is underspecified, which could turn the reported gains into leakage. read the letter →

arxiv 2502.06557 v1 pith:U3GJAJW6 submitted 2025-02-10 cs.IR

Yucheng Lu , Jiangxia Cao , Xu Kuan , Wei Cheng , Wei Jiang , Jiaming Zhang , Yang Shuang , Liu Zhaojie
show 1 more author
Liyin Hong
This is my paper · ORCID
classification cs.IR
keywords live-streamingrecommendationfutureinformationpredictiongenerativemodeltransformertime-seriesforecastingproductcategoryrankingstatisticsequences
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

Live-streaming content changes by the second, and valuable user actions—buying, tipping, staying long—happen well after the first impression. The paper claims that a recommender can be improved by generating, rather than only measuring, what a stream will do a few minutes ahead: a transformer forecasts the stream's next behavior statistics (clicks, sales, gifts, comments) and the category of the next product to be sold, and these forecasts are fed into the ranking model as features. On a large production platform, the authors report offline AUC gains of about +0.1% for click-through and conversion prediction over an already strong industrial baseline, and online A/B gains in revenue and user engagement. The authors also position the work as the first to attack live-streaming recommendation from a future-information prediction perspective.

What carries the argument

The mechanism is a pair of Transformer encoders with a shared design philosophy. The statistic-sequence branch applies reversible instance normalization (ReVIN) before and after an inverted Transformer (iTransformer) that encodes $S_{1:T-5}$ and outputs the next five steps $\hat{S}_{T-4:T}$; the product-sequence branch embeds product IDs and three category levels and uses a vanilla Transformer to predict the distribution over the next fine-grained category $\hat{c}^3_{T+1}$. The generated forecasts, along with the final sequence encodings, are flattened and injected into the ranking model through a stop-gradient wrapper that prevents gradients from flowing from the ranking loss into the generator. This setup lets the generator learn purely from forecast accuracy while the ranker uses the forecast as a feature.

What would settle it

A direct test would be to rerun the offline evaluation with the statistic-sequence features shifted one step later in time (so the model only sees strictly past aggregates) and check whether the AUC gains persist; a second check is to compute the correlation between the forecast values and the actual next-window statistics on a held-out sample, because if the correlation is high only when the windows overlap, the future-information claim is weakened.

Watch

Extended reading notes

Core claim

The central discovery is that future-information generation is a viable and effective input for live-streaming ranking. LiveForesighter maintains two sequence models: one over 30-second aggregated user-behavior statistics (exposure, audience, GMV, orders, gifts, comments, clicks) that is trained with mean-squared error to predict the next five steps, and one over the chronological product IDs and their three-level category codes that is trained with cross-entropy to predict the fine-grained category of the next product. At serving time, the model forecasts the next few statistic steps and the next product category, flattens the forecasts and sequence encodings, and concatenates them into the ranking model's representation of the live stream, using stop-gradient so the generator's loss is independent of the ranker. The paper argues that this let the ranker detect 'high-light moments' and anticipate future content, and it reports that the component improves offline CTR/CVR AUC and online GMV and infiltration metrics, with the effect growing as forecast accuracy grows.

Load-bearing premise

The account assumes that the statistic sequences are aligned in time with the exposure and label windows so that the 'future' values predicted are genuinely not yet observed; if the 30-second aggregation window already contains the label information the model is meant to forecast, the reported gains could come from label leakage rather than real anticipation.

Editorial extensions

If this is right

  • If the gains hold, live-streaming recommenders can add future-information forecasts as cheap features without changing the ranking architecture.
  • The stop-gradient design lets the generator be trained separately from the ranker, making the approach modular and easy to swap with better forecasters.
  • The correlation between forecast accuracy and ranking quality implies that improving the forecasters (longer horizons, better models) should directly improve recommendations.
  • The same future-generation idea applies to other time-sensitive recommendation surfaces, such as short-video feeds or e-commerce product sequences.

Reading between the lines

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

  • One implication the paper does not spell out is that the statistic-sequence forecasts may partly encode the same behaviors used as ranker labels, so the measured gains could overstate the value of true anticipation (this is the timing assumption highlighted in the weakest-assumption field).
  • A second extension: the product-category forecaster assumes a streamer's product style is stable; a natural stress test is whether the gains survive when a streamer abruptly changes product categories or when the model is evaluated on a cold-start stream with no history.
  • A third: the approach treats forecast features as flat inputs to the ranker; a more direct use would be to feed the forecast distribution (not just its mode) or to regularize the ranker with the forecast as an auxiliary task.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes LiveForesighter, a generative module for live-streaming recommendation at Kuaishou. It augments a large industrial ranking model with two auxiliary forecasting tasks: (1) forecasting future 30-second-aggregated live-streaming statistics such as audience, GMV, orders, gift value, comments, likes, and in-room clicks/carts, and (2) forecasting the fine-grained category of the next product to be sold in online-shopping live-streams. The forecasts and current-sequence encodings are fed into the ranking model with stop-gradient and trained with auxiliary supervised losses. The paper reports offline AUC/UAUC/GAUC gains on online-shopping and talent-show services and online A/B gains in GMV, infiltration, and related metrics, and states that the system has been deployed at Kuaishou. The main claimed contribution is that explicitly forecasting future live-streaming information improves ranking quality, and the paper states this is the first work to enhance live-streaming recommendation from a future-information prediction perspective.

Significance. If the empirical claims are valid, the paper demonstrates a practically useful paradigm for live-streaming recommendation: instead of relying only on current features, the ranker is given forecasted near-future statistics and product categories. The industrial setting, the deployment description, and the online A/B results give the work real-world relevance, and the two forecasting objectives are clearly specified as standalone supervised losses. The paper is honest about the need for stop-gradient and about the separate contributions of statistic and product sequences. However, the central claim depends on a temporal-separation assumption that is never stated: the relationship between the sequence time index T, the recommendation exposure time, and the label observation window is not defined. Because the statistic sequences aggregate exactly the behavior types that define the ranking labels, this omission creates a serious leakage risk. The evidence also lacks variance estimates, confidence intervals, and an ablation that isolates the forecast component from the current-sequence encoding. For these reasons, the contribution is promising but not yet established as stated.

major comments (4)
  1. [§2.2, Eqs. (2)–(3)] The paper never defines the time index T relative to the recommendation exposure time or to the label observation window used for CTR/CVR/long-view labels. The statistic sequences S include GMV, order, gift, click, and in-room click/cart counts — the same behavior types that constitute the labels. If the window S_{5:T} used for ranking overlaps with or extends into the label window, then both the current encoding E^{Stat}_{5:T} and the forecast bS_{T:T+5} can encode label information, and the reported gains in Tables 1–4 may reflect leakage rather than genuine future prediction. This concern is sharpened by Table 7, where interaction and in-live features alone give +0.19% CVR-AUC. The authors should specify the exact timing of T relative to exposure and label observation, and should verify that the model uses only information available at the moment of scoring.
  2. [§2.2 and §4.5] There is a direct inconsistency in the forecasting horizon. Equation (2) states 'we assume to predict next 5 steps,' and Eq. (3) constructs bS_{T:T+5}, but Section 4.5 states 'we only predict the next 3 steps for those statistic sequence in training and inference.' The same inconsistency affects the reported forecast-step analysis in Table 8, where Step-2 and Step-3 results are shown for a model whose stated equations use a 5-step horizon. The authors must reconcile the equations and the experimental setting, and report which horizon was used in the deployed and offline models.
  3. [§4.2, Tables 1–2] The ablations labeled '+ Statistic Sequences' and '+ Product Sequences' combine the current-sequence encoding and the forecast output, so they do not isolate the contribution of forecasting the future from the contribution of feeding current aggregate statistics into the ranker. A control that includes the current encoding but replaces the forecast with, for example, the latest observed value or mean value would directly test the paper's central claim that generating future information, rather than merely using current statistics, drives the gain. Tables 5 and 6 compare forecasting methods but always keep the current encoding in the model, so they do not resolve this confound.
  4. [§4.2–4.3, Tables 1–4] All offline and online results are reported as point estimates without variance, confidence intervals, or significance tests. Given that the headline offline gains are 0.03–0.15 AUC points and the online gains are 0.06–0.86% relative changes, the absence of statistical uncertainty makes it difficult to judge whether the improvements are stable. The claim in Section 4.2 that 'stable improvements of 0.05% in AUC' are significant is an assertion, not a demonstrated property; the authors should report per-experiment variance (e.g., across training runs or bootstrap over evaluation days) and, for the A/B tests, the number of users/exposures and confidence intervals or p-values.
minor comments (5)
  1. [§2.2, Eq. (2)] The notation S = {s1, ..., s_N} ∈ R^{N×T} is used both for the collection of statistic sequences and for the input tensor; this overload is confusing, especially because s_i is a sequence but S also denotes the tensor. Consider using different symbols for the sequence set and the matrix.
  2. [§2.3, Eq. (4)] Eq. (4) writes L_Product = Softmax(bc3_T, c3_T), but 'Softmax' is not a loss function; this should be a cross-entropy objective such as CE(bc3_T, c3_T). The current notation is imprecise.
  3. [§4.1] The paper reports that the baseline ranking models are 'huge MLP-based neural-networks' but gives no detail on their feature sets, training data volume, or training window. Since the paper claims to compare against a strong industrial baseline, a brief description of the baseline's inputs and training setup is needed for reproducibility.
  4. [§5.1] The claim that LiveForesighter is 'the first work to enhance live-streaming recommendation from future information prediction perspective' is likely too strong; several time-series and sequence-prediction methods have been applied to live-streaming contexts. Please soften the claim or provide a more thorough comparison.
  5. [§4.6, Figure 5] Figure 5 shows two example live-streaming traces. The figure would be easier to interpret if the caption described what the red, blue, and gray curves represent, and if the x-axis units and the time alignment between exposure and CVR were stated.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the forecasting objectives are standalone auxiliary losses with stop-gradient detachment, and the self-citations are background material, not load-bearing derivation.

full rationale

The derivation chain is self-contained with respect to the paper's central claim. The forecasting objectives in Eq. (2) and Eq. (4) are standalone supervised losses over observed statistic and product-category sequences; the ranking loss in Eq. (1) is a separate binary cross-entropy objective, and the forecast features are fed to the ranker through stop-gradient in Eqs. (3) and (5), so the ranker cannot retroactively shape the forecasts. No fitted parameter is renamed as a prediction: the Transformer forecasters are trained on future statistics and categories, and their outputs are used as covariates. Self-citations such as Moment&Cross [5], KuaiFormer [19], QARM [23], MARM [24], and HoME [37] appear only in Related Work or baseline descriptions and are not load-bearing mathematical premises. The paper's timing ambiguity — T is not anchored to exposure or label windows, and the statistic aggregates include behaviors of the same type as the labels — is a potential label-leakage or soundness concern, not an equivalence-by-construction; the text provides no equation or definition that makes the reported AUC gains a mathematical consequence of the forecasting loss. The training/inference horizon inconsistency (Eq. (2) says 5 steps; Section 4.5 says 3 steps) is a reporting inconsistency, not circularity. Overall, no significant circularity is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method rests on two domain observations, consistent style and high-light moment, and on standard Transformer implementations. There are no invented physical entities. The forecast horizon and the 30-second sampling interval are hand-chosen and are the main tunable quantities; the timing and leakage assumption is unstated and load-bearing.

free parameters (3)
  • Forecast horizon for statistic sequences = 3 steps (Eq. (2) writes 5 steps)
    Section 2.2 defines S_{1:T-5} to predict the next 5 steps, but Section 4.5 says only 3 steps are used. Table 8 shows step-2 gives higher CTR-AUC (+0.07%) than step-3 (+0.03%), so the chosen horizon is hand-selected and not the best on the reported metric.
  • Statistic sequence sampling interval = 30 seconds
    Section 2.2 collects statistic sequences every 30 seconds. This granularity is chosen by hand and is not ablated.
  • Product forecasting target granularity = Third-level category (c3) only
    Section 2.3 predicts only the finest product category, not the product ID, as a cost and efficiency trade-off. The 5000-class softmax head is a fitted layer and this choice constrains the forecast information.
assumptions (4)
  • domain assumption Live-streaming authors have a consistent style, so future products and behavior trends can be predicted from past content.
    Invoked in the Introduction and Section 2.3 to justify product sequence forecasting and to ground the high-light moment detection.
  • domain assumption A rising trend in users' positive behaviors identifies a high-light moment, and showing the stream during high-light moments improves user experience.
    Section 2.2 uses this observation to justify the statistic-sequence forecasting objective as a proxy for content quality.
  • standard math Standard Transformer, iTransformer, and RevIN implementations behave as described in their cited papers.
    The paper relies on cited models for sequence encoding and normalization without re-deriving them.
  • domain assumption No label leakage occurs between the statistic forecast window and the ranking label window.
    The paper never defines T relative to exposure time; the validity of the empirical gains depends on this implicit timing assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LiveForesighter: Generating Future Information for Live-Streaming Recommendations at Kuaishou." pith.science (2026). https://pith.science/paper/U3GJAJW6

@misc{pith2026250206557,
  author       = {Pith},
  title        = {Pith review of: LiveForesighter: Generating Future Information for Live-Streaming Recommendations at Kuaishou},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U3GJAJW6}},
  note         = {Machine review of arXiv:2502.06557}
}
read the original abstract

Live-streaming, as a new-generation media to connect users and authors, has attracted a lot of attention and experienced rapid growth in recent years. Compared with the content-static short-video recommendation, the live-streaming recommendation faces more challenges in giving our users a satisfactory experience: (1) Live-streaming content is dynamically ever-changing along time. (2) valuable behaviors (e.g., send digital-gift, buy products) always require users to watch for a long-time (>10 min). Combining the two attributes, here raising a challenging question for live-streaming recommendation: How to discover the live-streamings that the content user is interested in at the current moment, and further a period in the future?

Figures

Figures reproduced from arXiv: 2502.06557 by the authors.

Figure 1
Figure 1. Live-streaming high-lights could be reflected by the crowd of users behaviours. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The interpreting product sequence of a online-shopping live-streaming. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. An overview of model architecture of LiveForesighter. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: LiveForesighter Deployment. informationbc 3 𝑇 +1 , our ranking model not only determine whether users currently enjoy the live-streaming, but also predict whether they will continue to like its content in the near future. The overall framework of our LiveForesighter is…
Figure 5
Figure 5. Figure 5: Live-streaming exposure trend across time. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. KuaiLive-M3: A Multi-Modal, Multi-Domain, and Multi-Feedback Dataset for Live Streaming Recommendation

    cs.IR 2026-07 conditional novelty 6.0 of 10

    KuaiLive-M3 releases multi-domain Kuaishou logs, ~88M segment multi-modal embeddings, and 25k questionnaires, with benchmarks showing gains from cross-domain transfer, temporal modeling, and sparse explicit feedback.

  2. Towards Generalizable Safety in Crowd Navigation via Conformal Uncertainty Handling

    cs.RO 2025-08 unverdicted novelty 5.0 of 10

    A crowd navigation method augmenting reinforcement learning with conformal uncertainty estimates is claimed to cut collisions under distribution shift, but the manuscript body is an unrelated live streaming dataset paper.

Reference graph

Works this paper leans on

51 extracted references · 47 canonical work pages · cited by 2 Pith papers

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. Arxiv (2023)

    work page 2023

  2. [2]

    Keyu An, Qian Chen, Chong Deng, Zhihao Du, Changfeng Gao, Zhifu Gao, Yue Gu, Ting He, Hangrui Hu, Kai Hu, et al. 2024. Funaudiollm: Voice understanding and generation foundation models for natural interaction between humans and llms. Arxiv (2024)

    work page 2024

  3. [3]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report. Arxiv (2023)

    work page 2023

  4. [4]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. In Conference on Neural Information Processing Systems (NeurIPS)

    work page 2020

  5. [5]

    Jiangxia Cao, Shen Wang, Yue Li, Shenghui Wang, Jian Tang, Shiyao Wang, Shuang Yang, Zhaojie Liu, and Guorui Zhou. 2024. Moment&Cross: Next- Generation Real-Time Cross-Domain CTR Prediction for Live-Streaming Recom- mendation at Kuaishou. Arxiv (2024)

    work page 2024

  6. [6]

    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 ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)

    work page 2023

  7. [7]

    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 ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)

    work page 2023

  8. [8]

    Olivier Chapelle. 2014. Modeling delayed feedback in display advertising. In ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)

    work page 2014

Show all 51 references
  1. [9]

    Edgar Chen, Mark Ally, Eder Santana, and Saad Ali. 2022. Weighing dynamic availability and consumption for Twitch recommendations. Arxiv (2022)

    2022

  2. [10]

    Junyi Chen, Lu Chi, Bingyue Peng, and Zehuan Yuan. 2024. Hllm: Enhancing sequential recommendations via hierarchical large language models for item and user modeling. Arxiv (2024)

    2024

  3. [11]

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

  4. [12]

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

  5. [13]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep Neural Networks for YouTube Recommendations. In ACM Conference on Recommender Systems (RecSys)

    2016

  6. [14]

    Jiaxin Deng, Dong Shen, Shiyao Wang, Xiangyu Wu, Fan Yang, Guorui Zhou, and Gaofeng Meng. 2023. ContentCTR: Frame-level Live Streaming Click-Through Rate Prediction with Multimodal Transformer. Arxiv (2023)

    2023

  7. [15]

    Jiaxin Deng, Shiyao Wang, Yuchen Wang, Jiansong Qi, Liqin Zhao, Guorui Zhou, and Gaofeng Meng. 2024. MMBee: Live Streaming Gift-Sending Recommen- dations via Multi-Modal Fusion and Behaviour Expansion. In ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)

    2024

  8. [16]

    Jui-Ting Huang, Ashish Sharma, Shuying Sun, Li Xia, David Zhang, Philip Pronin, Janani Padmanabhan, Giuseppe Ottaviano, and Linjun Yang. 2020. Embedding- based retrieval in facebook search. In ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)

    2020

  9. [17]

    Taesung Kim, Jinhee Kim, Yunwon Tae, Cheonbok Park, Jang-Ho Choi, and Jaegul Choo. 2021. Reversible instance normalization for accurate time-series forecasting against distribution shift. In International Conference on Learning Representations (ICLR)

    2021

  10. [18]

    Fengqi Liang, Baigong Zheng, Liqin Zhao, Guorui Zhou, Qian Wang, and Yanan Niu. 2024. Ensure Timeliness and Accuracy: A Novel Sliding Window Data Stream Paradigm for Live Streaming Recommendation. Arxiv (2024)

    2024

  11. [19]

    Chi Liu, Jiangxia Cao, Rui Huang, Kai Zheng, Qiang Luo, Kun Gai, and Guorui Zhou. 2024. KuaiFormer: Transformer-Based Retrieval at Kuaishou.Arxiv (2024)

    2024

  12. [20]

    Qi Liu, Kai Zheng, Rui Huang, Wuchao Li, Kuo Cai, Yuan Chai, Yanan Niu, Yiqun Hui, Bing Han, Na Mou, et al . 2024. RecFlow: An Industrial Full Flow Recommendation Dataset. Arxiv (2024)

    2024

  13. [21]

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. 2023. itransformer: Inverted transformers are effective for time series forecasting. Arxiv (2023)

    2023

  14. [22]

    Yong Liu, Haoran Zhang, Chenyu Li, Xiangdong Huang, Jianmin Wang, and Mingsheng Long. 2024. Timer: Generative Pre-trained Transformers Are Large Time Series Models. In International Conference on Machine Learning (ICML)

    2024

  15. [23]

    Xinchen Luo, Jiangxia Cao, Tianyu Sun, Jinkai Yu, Rui Huang, Wei Yuan, Hezheng Lin, Yichen Zheng, Shiyao Wang, Qigen Hu, et al. 2024. QARM: Quantitative Alignment Multi-Modal Recommendation at Kuaishou. Arxiv (2024)

    2024

  16. [24]

    Xiao Lv, Jiangxia Cao, Shijie Guan, Xiaoyou Zhou, Zhiguang Qi, Yaqiang Zang, Ming Li, Ben Wang, Kun Gai, and Guorui Zhou. 2024. MARM: Unlocking the Future of Recommendation Systems through Memory Augmentation and Scalable Complexity. Arxiv (2024)

    2024

  17. [25]

    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 International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR)

    2018

  18. [26]

    Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2022. A time series is worth 64 words: Long-term forecasting with transformers. Arxiv (2022)

    2022

  19. [27]

    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. InACM International Conference on Information and Knowledge Managem...

    2020

  20. [28]

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Tran, Jonah Samost, et al

  21. [29]

    Jérémie Rappaz, Julian McAuley, and Karl Aberer. 2021. Recommendation on live-streaming platforms: Dynamic availability and repeat consumption. In ACM Conference on Recommender Systems (RecSys)

    2021

  22. [30]

    Steffen Rendle. 2010. Factorization machines. In IEEE International Conference on Data Mining (ICDM)

    2010

  23. [31]

    Tobias Schnabel, Adith Swaminathan, Ashudeep Singh, Navin Chandak, and Thorsten Joachims. 2016. Recommendations as treatments: Debiasing learning and evaluation. In International Conference on Machine Learning (ICML)

    2016

  24. [32]

    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 (2024)

    2024

  25. [33]

    I Sutskever. 2014. Sequence to Sequence Learning with Neural Networks. Arxiv (2014)

    2014

  26. [34]

    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 ACM Conference on Recommender Systems (RecSys)

    2020

  27. [35]

    Ye Tian, Ling Yang, Haotian Yang, Yuan Gao, Yufan Deng, Jingmin Chen, Xintao Wang, Zhaochen Yu, Xin Tao, Pengfei Wan, et al. 2024. VideoTetris: Towards Compositional Text-to-Video Generation. Arxiv (2024)

    2024

  28. [36]

    A Vaswani. 2017. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS)

    2017

  29. [37]

    Xu Wang, Jiangxia Cao, Zhiyi Fu, Kun Gai, and Guorui Zhou. 2025. HoME: Hierarchy of Multi-Gate Experts for Multi-Task Learning at Kuaishou. In ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)

    2025

  30. [38]

    Dinghao Xi, Liumin Tang, Runyu Chen, and Wei Xu. 2023. A multimodal time- series method for gifting prediction in live streaming platforms. Information Processing & Management (IPM) (2023)

    2023

  31. [39]

    Jing Yan, Liu Jiang, Jianfei Cui, Zhichen Zhao, Xingyan Bin, Feng Zhang, and Zuotao Liu. 2024. Trinity: Syncretizing Multi-/Long-tail/Long-term Interests All in One. Arxiv (2024)

    2024

  32. [40]

    Liu Yang, Fabian Paischer, Kaveh Hassani, Jiacheng Li, Shuai Shao, Zhang Gabriel Li, Yun He, Xue Feng, Nima Noorshams, Sem Park, et al. 2024. Unifying Generative and Dense Retrieval for Sequential Recommendation. Arxiv (2024)

    2024

  33. [41]

    Xiaoyong Yang, Yadong Zhu, Yi Zhang, Xiaobo Wang, and Quan Yuan. 2020. Large scale product graph construction for recommendation in e-commerce. Arxiv (2020)

    2020

  34. [42]

    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 (2024)

    2024

  35. [43]

    Chao Zhang, Shiwei Wu, Haoxin Zhang, Tong Xu, Yan Gao, Yao Hu, and En- hong Chen. 2024. NoteLLM: A Retrievable Large Language Model for Note Recommendation. In Companion Proceedings of the ACM on Web Conference

    2024

  36. [44]

    Yixin Zhang, Yong Liu, Hao Xiong, Yi Liu, Fuqiang Yu, Wei He, Yonghui Xu, Lizhen Cui, and Chunyan Miao. 2023. Cross-domain disentangled learning for e-commerce live streaming recommendation. In IEEE International Conference on Data Engineering (ICDE)

    2023

  37. [45]

    Yu Zhang and Qiang Yang. 2022. A Survey on Multi-Task Learning. IEEE Transactions on Knowledge and Data Engineering (TKDE) (2022)

    2022

  38. [46]

    Zhe Zhao, Lichan Hong, Li Wei, Jilin Chen, Aniruddh Nath, Shawn Andrews, Aditee Kumthekar, Maheswaran Sathiamoorthy, Xinyang Yi, and Ed Chi. 2019. Recommending what video to watch next: a multitask ranking system. In ACM Conference on Recommender Systems (RecSys)

    2019

  39. [47]

    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 Conference on Knowledge Discovery and Data Mining (KDD)

    2018

  40. [48]

    Han Zhu, Daqing Chang, Ziru Xu, Pengye Zhang, Xiang Li, Jie He, Han Li, Jian Xu, and Kun Gai. 2019. Joint optimization of tree-based index and deep model for recommender systems. In Advances in Neural Information Processing Systems (NeurIPS)

    2019

  41. [2016]

    In ACM Conference on Recommender Systems (RecSys Workshop)

    Wide & deep learning for recommender systems. In ACM Conference on Recommender Systems (RecSys Workshop)

  42. [2021]

    Arxiv (2021)

    End-to-end user behavior retrieval in click-through rateprediction model. Arxiv (2021)

    2021

  43. [2023]

    Recommender systems with generative retrieval

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.