Pith. sign in

REVIEW 3 major objections 7 minor 56 references

A public live-streaming dataset that finally joins evolving multi-modal content, short-video behavior, and explicit user feedback in one benchmark.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-30 17:00 UTC pith:I2X6XMXD

load-bearing objection Solid industrial dataset release that fills a real public-data gap; one interpretive claim on “evolving content” is softer than the Abstract sells. the 3 major comments →

arxiv 2607.24862 v1 pith:I2X6XMXD submitted 2026-07-26 cs.IR

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

Ke Guo , Changle Qu , Jiayaqi Cheng , Xiao Zhang , Shijun Wang , Xiaoyu Zhang , Xueliang Wang , Le Zhang
show 2 more authors
Lantao Hu Jun Xu
This is my paper
classification cs.IR
keywords DatasetLive StreamingCross-Domain RecommendationMulti-Modal EmbeddingsHighlight PredictionQuestionnaire FeedbackSequential Recommendation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Public live-streaming recommendation research has been stuck with datasets that miss three real-world ingredients: content that changes minute by minute, users who also watch short videos, and any direct signal of whether people actually want a streamer recommended. This paper releases KuaiLive-M3, a Kuaishou-scale collection of about 22,000 users with tens of millions of live interactions, over a hundred million short-video interactions, roughly 88 million timestamped segment embeddings, and 25,403 questionnaire answers about recommendation frequency. On top of the release it defines three tasks—cross-domain recommendation, next-segment highlight prediction, and questionnaire-enhanced ranking—and shows that specialized transfer models, temporal models, and sparse explicit feedback each move the needle where generic baselines do not. A sympathetic reader cares because the missing pieces are exactly what industrial live systems already use, so a public benchmark that includes them lets the field test ideas that previously lived only inside companies.

Core claim

KuaiLive-M3 is a challenging, realistic public benchmark for live streaming recommendation: its multi-modal segment embeddings, short-video–live cross-domain logs, and questionnaire feedback enable tasks prior public datasets cannot support, and representative baselines show that modeling temporally evolving content, transferring preferences across domains, and bridging implicit versus explicit feedback each matter.

What carries the argument

KuaiLive-M3 itself—the joint release of timestamped multi-behavior logs across short video and live domains, PCA-reduced segment- and room-level multi-modal embeddings, and streamer-frequency questionnaires—plus the three standardized tasks (cross-domain ranking of streamers, next-segment highlight scoring via retention and engagement density, and questionnaire-augmented sequential recommendation) that turn those signals into measurable benchmarks.

Load-bearing premise

The industrial multi-modal embeddings and the randomly thinned short-video logs are assumed to be faithful enough stand-ins for real evolving content and cross-domain density that conclusions drawn on them still transfer to live systems.

What would settle it

Re-run the three benchmarks with independently extracted open multi-modal features on the same rooms and with full (non-downsampled) short-video histories for the same users; if MGCCDR-style transfer, sequential highlight gains over MLP, and questionnaire lifts disappear or reverse, the claimed research utility of the released signals fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Cross-domain live recommenders can be compared publicly using shared authors and short-video play as transfer bridges rather than only overlapping users.
  • Highlight prediction can be trained and scored as next-segment forecasting from past segment embeddings without peeking at future content.
  • Sparse questionnaire labels become a standard second signal for testing whether explicit satisfaction improves over play-only sequential models.
  • Staytime prediction, generative semantic-ID recommendation over evolving rooms, and LLM preference simulation gain a common public substrate.
  • Cold-start long-tail rooms and front-loaded viewer arrival become measurable design targets rather than anecdotes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The questionnaire sparsity result (tens of millions of plays versus ~25k answers) implies future work must treat explicit feedback as a rare teacher signal, not as a dense second behavior channel.
  • Author overlap as a structural bridge suggests heterogeneous graph or multi-graph designs will keep outperforming pure user-overlap CDR on this platform family.
  • If segment embeddings are only kept when some user interacted, highlight models may systematically under-represent silent early segments—worth a controlled ablation on rooms with denser segment coverage.
  • The same multi-domain logs could stress-test whether generative recommenders need dynamic semantic IDs that change as a room’s content drifts mid-broadcast.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The paper introduces KuaiLive-M3, a dataset collected from Kuaishou covering 21,938 sampled users over four weeks, with ~35.8M live-streaming and ~111M short-video interactions, ~88M timestamped segment-level MLLM embeddings (plus room- and video-level embeddings), and 25,403 questionnaire responses on streamer recommendation preferences. The authors argue that prior public live-streaming datasets lack (i) temporally evolving multi-modal content, (ii) cross-domain short-video↔live behaviors, and (iii) explicit feedback, and that KuaiLive-M3 fills all three gaps. They benchmark three tasks — cross-domain recommendation (MGCCDR best; LightGCN second), next-segment highlight prediction (sequential models ≫ content-only MLP), and questionnaire-enhanced sequential recommendation (small but consistent gains from questionnaire-augmented variants) — and conclude that temporal content modeling, cross-domain transfer, and implicit–explicit feedback bridging each matter. Data and code are publicly released under CC BY-NC-SA 4.0.

Significance. If the resource is as described, this is a substantial contribution to a field where public data is genuinely scarce. The combination of timestamped segment-level multi-modal embeddings (88M, enabling temporal content modeling), joint short-video/live logs with shared user and author identities (17.6% author overlap, §4.2.2), and questionnaire-based explicit feedback is not available in any prior public dataset (Table 1's comparison is accurate as far as I can verify). The construction is documented with unusual specificity: label formulas (Eqs. 1–3), percentile thresholding, 5-core filtering, chronological splits, and hyperparameter search grids (§5.1.3, §5.2.3, §5.3.3) are all stated, and code plus data are released with a clear license and an ethics/anonymization statement (§3.1.5–3.1.6). The honest reporting of negative or mixed results — LightGCN beating most CDR methods, SAQRec underperforming (§5.3.4) — increases confidence in the benchmark tables. The limitations (non-response bias, short-video downsampling, non-public MLLM) are candidly disclosed in §7. The empirical conclusions, however, are currently supported by benchmark designs that leave important alternative explain

major comments (3)
  1. [§5.2, Eqs. (1)–(3), Table 5] The conclusion that 'temporal dependencies among previously observed segments are essential' rests on the GRU-vs-MLP gap (Kendall's τ 0.507 vs 0.147; mAP 0.778 vs 0.515). But the supervision label has a strong built-in positional trend: in Eq. (1), N_entered is cumulative over the room's lifetime while N_stay shrinks as viewers leave, so LVTR mechanically decays with segment position — consistent with the monotonically decaying concurrent-viewer curve in Fig. 2(d). Within-room min-max normalization rescales but does not remove this trend, so y (Eq. 3) is substantially a smooth function of relative segment position. A sequential model can plausibly decode stream age from temporally correlated adjacent embeddings and extrapolate the trend, while the single-segment MLP cannot. The benchmark includes no content-free control: no persistence baseline (predict y_{k+1} = y_k), no decay-fit basel
  2. [§5.3.1, Table 6] The absolute metric values are implausibly high for streamer recommendation — HR@1 between 0.81 and 0.87, NDCG@10 ≈ 0.90–0.93. With 21,938 users, millions of authors, and 60.4% of rooms watched by a single user (Fig. 2(c)), ranking one held-out positive against 99 randomly sampled never-interacted streamers is close to trivial: almost any popularity signal separates the positive from random long-tail negatives. The reported metric ceiling therefore mostly reflects the easy-negative protocol, and the paper's central claim for this task — that questionnaire feedback improves recommendation — rests on relative deltas of 0.23%–1.18% (§5.3.4) reported with no variance estimates, no repeated runs with different negative samples, and no significance testing. At this ceiling, such deltas are within plausible seed/sampling noise. The authors should (a) report mean ± std over multiple negative-sam
  3. [§5.2.1, Eqs. (1)–(3)] The highlight labels are constructed solely from the play/like/comment logs of the 21,938 sampled users, but the benchmark selects the 10,000 rooms with the most playing interactions — precisely the rooms whose true audiences are far larger than the sampled cohort (per Fig. 2(c), popular rooms have hundreds to thousands of viewers, of which only a small fraction are sampled users). LVTR and ED (Eqs. 1–2) are therefore computed over a sparse, possibly unrepresentative sub-population of each room's audience, and segments with zero sampled-user events are excluded from embedding release (§3.1.3), which censors the label series further. The paper should quantify this: report the distribution of sampled-viewer counts per segment in the benchmark rooms, and ideally show label stability under cohort subsampling (e.g., correlation of y computed from disjoint halves of the sampled viewers). If pe
minor comments (7)
  1. [§3.1.2] §3.1.2 states that short-video interactions are 'randomly sample[d]' for scale but never reports the sampling rate or whether sampling is uniform over users, time, or both. Since §4.2.1's 3:1 density ratio and the CDR benchmark depend on this, the rate and scheme should be stated (cf. Limitations §7).
  2. [Figure 2(d), §4.1.3] Fig. 2(d) caption says 'average viewer arrival distribution' while the axis and main text describe concurrent viewers over normalized stream lifetime; these are different quantities. Please make terminology consistent.
  3. [Table 3, §3.1.3] Table 3 lists 6,502,107 room-level embeddings for 6,564,013 rooms; the ~62k gap (rooms with no retained segments) should be noted explicitly, as should the analogous gap for videos (5,498,631 embeddings vs 6,741,159 videos).
  4. [§5.3.1] §5.3.1 treats both 'Recommend whenever the streamer goes live' and 'Recommend occasionally' as positive, conflating two distinct preference intensities. A sensitivity analysis separating the two (or a brief justification) would strengthen the questionnaire benchmark, especially given the open-ended Q2 responses that could support graded labels.
  5. [Tables 4–5] Tables 4 and 5 report single-run point estimates; given the CDR result (LightGCN second-best, most CDR methods below single-domain baselines) is somewhat counterintuitive and cited as evidence of task difficulty, std over seeds would help readers calibrate the gaps.
  6. [§3.1.3 vs Appendix A] §3.1.3 says segments are 'uniformly segmented clips' while Appendix A and §5.2.1 describe variable-length segments delimited by embedding timestamps; please reconcile the terminology.
  7. [§3.1.3, §7] The reliance on a non-public industrial MLLM for all embeddings (§3.1.3) is acknowledged in §7, but the paper could additionally report basic embedding-quality diagnostics (e.g., a probe task such as predicting room/video category from embeddings) so users can gauge what the released vectors capture.

Circularity Check

0 steps flagged

No circular derivation: dataset/benchmark paper with standard held-out supervision, not self-definitional predictions.

full rationale

KuaiLive-M3 is a dataset-and-benchmark paper. Its load-bearing claims are empirical (baselines on chronological splits for CDR, next-segment highlight scoring, and questionnaire-augmented ranking), not first-principles derivations. Highlight labels (Eqs. 1–3) are explicit behavioral definitions used as training targets; models map content embeddings to held-out next-segment scores—standard supervised learning, not equating a fitted input to a claimed prediction. Cross-domain and questionnaire results likewise train on interaction/feedback logs and evaluate on held-out rankings. Self-citations (KuaiLive, MGCCDR, related Kuaishou work) supply prior context and one specialized baseline; none import a uniqueness theorem or force the reported metrics by construction. Concerns that LVTR/ED may be positionally autocorrelated affect experimental interpretation of “content evolution,” not circularity of a derivation chain. No step reduces a claimed prediction to its inputs by definition.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

As a dataset-and-benchmark paper, load-bearing commitments are domain conventions for recsys evaluation, platform-specific logging definitions, and modeling choices for labels—not physical axioms. Free parameters are engineering choices in label construction and filtering that affect reported metrics. Invented entities are none beyond the dataset name itself as a packaged resource.

free parameters (5)
  • highlight score mix weights (0.6 LVTR + 0.4 ED) = 0.6 / 0.4
    Continuous segment supervision is a hand-chosen convex combination following prior work and 'empirical tuning' (§5.2); different weights would relabel highlights.
  • highlight binary threshold (70th percentile within room) = 70th percentile
    Binary mAP labels depend on this percentile cutoff (§5.2).
  • short-video positive watch-progress threshold = 10%
    CDR positives require >10% watch progress in short video (§5.1.1).
  • PCA embedding dimensions (128 segment/video, 64 room) = 128 / 64
    Released multi-modal features are PCA-reduced industrial MLLM outputs (§3.1.3); dimension is a design choice affecting all content-aware tasks.
  • 5-core filtering and chronological 8:1:1 or leave-one-out splits = 5-core; 8:1:1 or LOO
    Standard but consequential preprocessing that defines who/what remains in each benchmark (§5.1.1, §5.3.1).
axioms (5)
  • domain assumption Implicit play/engagement logs and questionnaire options are valid proxies for preference suitable for ranking metrics (Recall, NDCG, HR, MRR, Kendall τ, mAP).
    Underpins all three benchmarks in §5; standard in recsys but not proved here.
  • domain assumption Authors treated as recommendation items in live CDR/questionnaire tasks adequately represent live-room recommendation.
    Stated in §5.1.1 and §5.3.1; collapses room-level dynamics into streamer-level targets.
  • domain assumption Next-segment highlight labels from concurrent retention and like/comment density measure 'engaging moments' useful for proactive recommendation.
    Eqs. 1–3 and task framing in §5.2; alternative engagement definitions could change rankings.
  • domain assumption Chronological splits without leakage and random negative sampling (99 negs in questionnaire task) yield fair offline evaluation of online live systems.
    Evaluation protocols §5.1–5.3; common practice, still an assumption about external validity.
  • ad hoc to paper Uniform random questionnaire distribution among eligible users yields usable explicit feedback despite possible non-response bias.
    Collection §3.1.4; Limitations §7 admits selection bias risk.
invented entities (1)
  • KuaiLive-M3 dataset package (joined multi-domain logs, PCA MLLM embeddings, questionnaire table, task splits) independent evidence
    purpose: Serve as public multi-modal, multi-domain, multi-feedback benchmark resource for live streaming recommendation.
    The named artifact is the paper's primary contribution; it is a curated release, not a new physical entity. independent_evidence is true insofar as the files are publicly downloadable and inspectable.

pith-pipeline@v1.2.0-grok45-kimik3 · 23611 in / 3876 out tokens · 83364 ms · 2026-07-30T17:00:19.488357+00:00 · methodology

0 comments
read the original abstract

Existing public live streaming datasets suffer from three major limitations: they provide limited access to temporally evolving multimodal live content, overlook users' cross-domain interactions between short videos and live streams, and contain only implicit behavioral signals without explicit feedback that captures users' perceived content quality and satisfaction. These limitations prevent existing benchmarks from faithfully reflecting real-world live streaming scenarios and hinder comprehensive research on live streaming recommendation. To address these limitations, we introduce KuaiLive-M3, a multi-modal, multi-domain, and multi-feedback dataset for live streaming recommendation, collected from Kuaishou, a leading live streaming and short video platform in China. KuaiLive-M3 covers 21,938 users and contains 35 million live streaming interactions and 111 million short video interactions, with fine-grained timestamps and diverse user behaviors. It further provides approximately 88 million timestamped segment-level multi-modal embeddings that capture the temporal evolution of live streaming content, as well as 25,403 questionnaire-based feedback records that bridge implicit user behaviors and explicit user preferences. Based on these unique signals, we establish benchmarks for cross-domain recommendation, live stream highlight prediction, and questionnaire-enhanced recommendation. Extensive experiments with representative baselines demonstrate that KuaiLive-M3 provides a challenging and realistic benchmark for future live streaming recommendation research. The results further highlight the importance of modeling temporally evolving content, transferring user preferences across domains, and bridging the gap between implicit behaviors and explicit user feedback. The dataset and benchmark code are publicly available at https://imgkkk574.github.io/KuaiLive-M3/.

Figures

Figures reproduced from arXiv: 2607.24862 by Changle Qu, Jiayaqi Cheng, Jun Xu, Ke Guo, Lantao Hu, Le Zhang, Shijun Wang, Xiaoyu Zhang, Xiao Zhang, Xueliang Wang.

Figure 1
Figure 1. Figure 1: Overview of the KuaiLive-M3 Dataset. public datasets primarily focus on the live streaming domain alone and rely predominantly on implicit interaction signals [25, 28, 40]. In particular, KuaiLive [25] advances this line of research by pro￾viding realistic, timestamped sequences of user interactions in an interactive live streaming environment. However, it does not cap￾ture users’ behaviors in the closely … view at source ↗
Figure 2
Figure 2. Figure 2: Analysis of interaction patterns in the live streaming domain. (a) shows the distribution of live room durations. (b) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

56 extracted references · 12 linked inside Pith

  1. [1]

    Jiangxia Cao, Xixun Lin, Xin Cong, Jing Ya, Tingwen Liu, and Bin Wang. 2022. Disencdr: Learning disentangled representations for cross-domain recommenda- tion. InProceedings of the 45th International ACM SIGIR conference on research and development in information retrieval. 267–277

    2022

  2. [2]

    Sunhao Dai, Ninglu Shao, Haiyuan Zhao, Weijie Yu, Zihua Si, Chen Xu, Zhongx- iang Sun, Xiao Zhang, and Jun Xu. 2023. Uncovering ChatGPT’s Capabilities in Recommender Systems. InProceedings of the 17th ACM Conference on Rec- ommender Systems. Association for Computing Machinery, New York, NY, USA, 1126–1132. doi:10.1145/3604915.3610646

    arXiv 2023

  3. [3]

    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 preprint arXiv:2306.14392 (2023)

    Pith/arXiv arXiv 2023

  4. [4]

    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)

    Pith/arXiv arXiv 2025

  5. [5]

    Jiaxin Deng, Shiyao Wang, Dong Shen, Liqin Zhao, Fan Yang, Guorui Zhou, and Gaofeng Meng. 2024. A multimodal transformer for live streaming highlight prediction. In2024 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 1–6

    2024

  6. [6]

    Jiaxin Deng, Shiyao Wang, Yuchen Wang, Jiansong Qi, Liqin Zhao, Guorui Zhou, and Gaofeng Meng. 2024. Mmbee: Live streaming gift-sending recommendations via multi-modal fusion and behaviour expansion. InProceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining. 4896–4905

    2024

  7. [7]

    Sheng Gao, Hao Luo, Da Chen, Shantao Li, Patrick Gallinari, and Jun Guo. 2013. Cross-domain recommendation via cluster-level latent factor model. InJoint European conference on machine learning and knowledge discovery in databases. Springer, 161–176

    2013

  8. [8]

    Yulong Gu, Zhuoye Ding, Shuaiqiang Wang, Lixin Zou, Yiding Liu, and Dawei Yin

  9. [9]

    Ke Guo, Changle Qu, Xiao Zhang, Liqin Zhao, Shijun Wang, Yanan Niu, and Jun Xu. 2026. Room Matters: Dynamic Room-level Collaboration Information Modeling for Live Streaming Recommendation. InProceedings of the ACM Web Conference 2026. 6045–6056

    2026

  10. [10]

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

    2020

  11. [11]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. InProceedings of the 26th international conference on world wide web. 173–182

    2017

  12. [12]

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

  13. [13]

    Yupeng Hou, Jiacheng Li, Xiangjun Fu, Zhankui He, An Yan, Xiusi Chen, and Julian McAuley. 2026. Bridging language and items for retrieval and recommen- dation: Benchmarking LLMs as semantic encoders. InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 3251–3265

    2026

  14. [14]

    Guangneng Hu, Yu Zhang, and Qiang Yang. 2018. CoNet: Collaborative Cross Networks for Cross-Domain Recommendation. InProceedings of the 27th ACM International Conference on Information and Knowledge Management(Torino, Italy)(CIKM ’18). Association for Computing Machinery, New York, NY, USA, 667–676. doi:10.1145/3269206.3271684

    arXiv 2018

  15. [15]

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

    2018

  16. [16]

    Jing Li, Pengjie Ren, Zhumin Chen, Zhaochun Ren, Tao Lian, and Jun Ma. 2017. Neural attentive session-based recommendation. InProceedings of the 2017 ACM on Conference on Information and Knowledge Management. 1419–1428

    2017

  17. [17]

    Xiaodong Li, Ruochen Yang, Shuang Wen, Shen Wang, Yueyang Liu, Guoquan Wang, Weisong Hu, Qiang Luo, Jiawei Sheng, Tingwen Liu, et al. 2026. FARM: Frequency-Aware Model for Cross-Domain Live-Streaming Recommendation. InProceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1. 2302–2311

    2026

  18. [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 preprint arXiv:2402.14399 (2024)

    Pith/arXiv arXiv 2024

  19. [19]

    Meng Liu, Jianjun Li, Guohui Li, and Peng Pan. 2020. Cross Domain Recommen- dation via Bi-directional Transfer Graph Collaborative Filtering Networks. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management(Virtual Event, Ireland)(CIKM ’20). Association for Computing Machinery, New York, NY, USA, 885–894. doi:10.1145/...

    arXiv 2020

  20. [20]

    Yueyang Liu, Jiangxia Cao, Shen Wang, Shuang Wen, Xiang Chen, Xiangyu Wu, Shuang Yang, Zhaojie Liu, Kun Gai, and Guorui Zhou. 2025. LLM-Alignment Live-Streaming Recommendation.arXiv preprint arXiv:2504.05217(2025)

    Pith/arXiv arXiv 2025

  21. [21]

    Zhanyu Liu, Shiyao Wang, Xingmei Wang, Rongzhou Zhang, Jiaxin Deng, Honghui Bao, Jinghao Zhang, Wuchao Li, Pengfei Zheng, Xiangyu Wu, et al

  22. [22]

    Yucheng Lu, Jiangxia Cao, Xu Kuan, Wei Cheng, Wei Jiang, Jiaming Zhang, Yang Shuang, Liu Zhaojie, and Liyin Hong. 2025. LiveForesighter: Generating Future Information for Live-Streaming Recommendations at Kuaishou.arXiv preprint arXiv:2502.06557(2025)

    Pith/arXiv arXiv 2025

  23. [23]

    Chen Ma, Peng Kang, and Xue Liu. 2019. Hierarchical gating networks for sequential recommendation. InProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining. 825–833

    2019

  24. [24]

    Tong Man, Huawei Shen, Xiaolong Jin, and Xueqi Cheng. 2017. Cross-Domain Recommendation: An Embedding and Mapping Approach. InProceedings of the Twenty-Sixth International Joint Conference on Artificial Intelligence, IJCAI-17. 2464–2470. doi:10.24963/ijcai.2017/343

    doi:10.24963/ijcai.2017/343 2017

  25. [25]

    Changle Qu, Sunhao Dai, Ke Guo, Xiao Zhang, Liqin Zhao, Shijun Wang, Yannan Niu, Lantao Hu, Han Li, and Jun Xu. 2025. KuaiLive: A Real-time Interactive Dataset for Live Streaming Recommendation.arXiv preprint arXiv:2508.05633 (2025)

    Pith/arXiv arXiv 2025

  26. [26]

    Changle Qu, Liqin Zhao, Yanan Niu, Xiao Zhang, and Jun Xu. 2025. Bridging Short Videos and Streamers with Multi-Graph Contrastive Learning for Live Streaming Recommendation. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2059–2069

    2025

  27. [27]

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

  28. [28]

    Jérémie Rappaz, Julian McAuley, and Karl Aberer. 2021. Recommendation on live-streaming platforms: Dynamic availability and repeat consumption. InPro- ceedings of the 15th ACM conference on recommender systems. 390–399

    2021

  29. [29]

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme

  30. [30]

    Teng Shi, Zhaoheng Li, Yuanhang Qu, Yi Liu, Lixiang Lai, and Yuning Jiang. 2026. SSRLive: Live Streaming Recommendation with Dynamic Semantic ID.arXiv preprint arXiv:2606.06970(2026)

    Pith/arXiv arXiv 2026

  31. [31]

    Singh and Geoffrey J

    Ajit P. Singh and Geoffrey J. Gordon. 2008. Relational learning via collective matrix factorization. InProceedings of the 14th ACM SIGKDD International Con- ference on Knowledge Discovery and Data Mining(Las Vegas, Nevada, USA) (KDD ’08). Association for Computing Machinery, New York, NY, USA, 650–658. doi:10.1145/1401890.1401969

    arXiv 2008

  32. [32]

    Jiaxi Tang and Ke Wang. 2018. Personalized top-n sequential recommenda- tion via convolutional sequence embedding. InProceedings of the eleventh ACM international conference on web search and data mining. 565–573

    2018

  33. [33]

    OneRec Team, Biao Yang, Boyang Ding, Chenglong Chu, Dunju Zang, Fei Pan, Han Li, Hao Jiang, Honghui Bao, Huanjie Wang, et al. 2026. OneReason Technical Report.arXiv preprint arXiv:2606.06260(2026)

    Pith/arXiv arXiv 2026

  34. [34]

    Shen Wang, Yusheng Huang, Ruochen Yang, Shuang Wen, Pengbo Xu, Jiangxia Cao, Yueyang Liu, Kuo Cai, Chengcheng Guo, Shiyao Wang, et al. 2026. OneLive: Dynamically Unified Generative Framework for Live-Streaming Recommenda- tion.arXiv preprint arXiv:2602.08612(2026)

    arXiv 2026

  35. [35]

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

    2019

  36. [36]

    Chuhan Wu, Fangzhao Wu, Tao Qi, Qi Liu, Xuan Tian, Jie Li, Wei He, Yongfeng Huang, and Xing Xie. 2022. Feedrec: News feed recommendation with various user feedbacks. InProceedings of the ACM Web Conference 2022. 2088–2097

    2022

  37. [37]

    Dinghao Xi, Liumin Tang, Runyu Chen, and Wei Xu. 2023. A multimodal time- series method for gifting prediction in live streaming platforms.Information Processing & Management60, 3 (2023), 103254

    2023

  38. [38]

    Ruobing Xie, Cheng Ling, Yalong Wang, Rui Wang, Feng Xia, and Leyu Lin. 2021. Deep feedback network for recommendation. InProceedings of the twenty-ninth international conference on international joint conferences on artificial intelligence. 2519–2525

    2021

  39. [39]

    Huan Yan, Xiangning Chen, Chen Gao, Yong Li, and Depeng Jin. 2019. Deepapf: Deep attentive probabilistic factorization for multi-site video recommendation. TC2, 130 (2019), 17–883

    2019

  40. [40]

    Sanshi Yu, Zhuoxuan Jiang, Dong-Dong Chen, Shanshan Feng, Dongsheng Li, Qi Liu, and Jinfeng Yi. 2021. Leveraging tripartite interaction information from live stream e-commerce for improving product recommendation. InProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 3886–3894

    2021

  41. [41]

    Jizhi Zhang, Keqin Bao, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2024. Large language models for recommendation: progresses and future directions. InCompanion Proceedings of the ACM Web Conference 2024. 1268–1271. Conference’17, July 2017, Washington, DC, USA Ke Guo et al

    2024

  42. [42]

    Kepu Zhang, Teng Shi, Sunhao Dai, Xiao Zhang, Yinfeng Li, Jing Lu, Xiaoxue Zang, Yang Song, and Jun Xu. 2024. Saqrec: Aligning recommender systems to user satisfaction via questionnaire feedback. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management. 3165–3175

    2024

  43. [43]

    Le Zhang, Xiaolan Zhu, Yuchen Wang, Shilong Kang, Jiaqi Xue, Xiaoyu Zhang, Xiang Chen, Yalong Guan, Xiangyu Wu, Shijun Wang, et al. 2026. Bridging Short Videos and Live Streams: Reasoning-Guided Multimodal LLMs for Cross-Domain Representation Learning.arXiv preprint arXiv:2606.04448(2026)

    Pith/arXiv arXiv 2026

  44. [44]

    Haiyuan Zhao, Changshuo Zhang, Yang Wang, Hao Wang, Zhen Ouyang, Bin Yuan, Qinglei Wang, and Zuotao Liu. 2025. Towards Unbiased and Real-Time Staytime Prediction for Live Streaming Recommendation. InProceedings of the 34th ACM International Conference on Information and Knowledge Management. Association for Computing Machinery, New York, NY, USA, 6293–630...

    arXiv 2025

  45. [45]

    Wayne Xin Zhao, Yupeng Hou, Xingyu Pan, Chen Yang, Zeyu Zhang, Zihan Lin, Jingsen Zhang, Shuqing Bian, Jiakai Tang, Wenqi Sun, et al. 2022. Recbole 2.0: Towards a more up-to-date recommendation library. InProceedings of the 31st ACM international conference on information & knowledge management. 4722– 4726

    2022

  46. [46]

    Wayne Xin Zhao, Shanlei Mu, Yupeng Hou, Zihan Lin, Yushuo Chen, Xingyu Pan, Kaiyuan Li, Yujie Lu, Hui Wang, Changxin Tian, et al. 2021. Recbole: Towards a unified, comprehensive and efficient framework for recommendation algorithms. Inproceedings of the 30th acm international conference on information & knowledge management. 4653–4664

    2021

  47. [47]

    Yang Zhao, Xuan Lin, Wenqiang Xu, Maozong Zheng, Zhengyong Liu, and Zhou Zhao. 2022. Antpivot: Livestream highlight detection via hierarchical attention mechanism.arXiv preprint arXiv:2206.04888(2022)

    Pith/arXiv arXiv 2022

  48. [48]

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

    2022

  49. [49]

    Feng Zhu, Chaochao Chen, Yan Wang, Guanfeng Liu, and Xiaolin Zheng. 2019. DTCDR: A Framework for Dual-Target Cross-Domain Recommendation. InPro- ceedings of the 28th ACM International Conference on Information and Knowledge Management(Beijing, China)(CIKM ’19). Association for Computing Machinery, New York, NY, USA, 1533–1542. doi:10.1145/3357384.3357992

    arXiv 2019

  50. [50]

    Feng Zhu, Yan Wang, Chaochao Chen, Guanfeng Liu, Mehmet Orgun, and Jia Wu

  51. [2012]

    BPR: Bayesian personalized ranking from implicit feedback.arXiv preprint arXiv:1205.2618(2012)

    Pith/arXiv arXiv 2012

  52. [2015]

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

    Pith/arXiv arXiv 2015

  53. [2018]

    InProceedings of the 27th International Joint Conference on Artificial Intelligence

    A deep framework for cross-domain and cross-system recommendations. InProceedings of the 27th International Joint Conference on Artificial Intelligence. 3711–3717. A Dataset Schema and Field Descriptions KuaiLive-M3 consists of 18 tables and collections covering shared entities, live streaming activities, and short video activities. Tables 7 and 8 summari...

    2017

  54. [2020]

    InProceedings of the 29th ACM international conference on information & knowledge management

    Deep multifaceted transformers for multi-objective ranking in large-scale e-commerce recommender systems. InProceedings of the 29th ACM international conference on information & knowledge management. 2493–2500

  55. [2023]

    Recommender systems with generative retrieval.Advances in Neural Information Processing Systems36 (2023), 10299–10315

    2023

  56. [2025]

    Onerec-think: In-text reasoning for generative recommendation.arXiv preprint arXiv:2510.11639(2025)

    arXiv 2025