Pith. sign in

REVIEW 4 major objections 6 minor 42 references

Describe What You See with Multimodal Large Language Models to Enhance Video Recommendations

T0 review · 4 major / 6 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 The paper claims that replacing raw video and audio features with MLLM-generated natural-language captions consistently improves video recommendation, with up to roughly 60% relative HR@10 gains on MicroLens-100K.

desk verdict Useful idea and clean pipeline, but the headline claims outrun the experiments: the video baseline is handicapped and the joint-caption result is never actually tested. read the letter →

arxiv 2508.09789 v1 pith:BYHYKNPC submitted 2025-08-13 cs.IR cs.CV

Marco De Nadai , Andreas Damianou , Mounia Lalmas This is my paper
classification cs.IRcs.CV
keywords videorecommendationmultimodallargelanguagemodelscontentcaptionszero-shotpromptingMicroLens-100Ktwo-towermodelsequentialaudiounderstanding
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

This paper tries to establish that the semantic gap in video recommendation can be closed by a cheap, model-agnostic step: prompt a frozen multimodal large language model (MLLM) to describe each clip in natural language, encode those descriptions with a text encoder, and feed them into standard recommenders. On the MicroLens-100K benchmark, this consistently outperforms conventional video, audio, and metadata features in a two-tower model and in SASRec, with audio captions alone giving a roughly 60% relative HR@10 increase. The authors argue that captions capture intent, humour, parody, and world knowledge that pixel and spectrogram features miss, and that because the MLLM is frozen and no finetuning is required, the pipeline is a low-barrier upgrade to existing systems. The result matters because short-form video platforms rely on exactly these resonance signals, and current encoders are blind to them.

What carries the argument

The load-bearing mechanism is the zero-finetuning MLLM captioning step: a frozen multimodal model compresses a minute-scale clip into a concise, knowledge-grounded sentence (e.g. a superhero parody with slapstick fights and orchestral stabs). This text is encoded with a single text encoder, so any recommender that consumes item embeddings can ingest it. The pipeline separates video and audio: Qwen-VL handles frames and on-screen text, while Whisper transcripts plus Qwen-Audio descriptions handle the soundtrack. The caption collapses long-horizon temporal content into one coherent representation, bypassing the fixed-frame limitations of video encoders like VideoMAE.

What would settle it

Give VideoMAE more than 16 frames, e.g. sample uniformly throughout the full clip rather than the first 30 seconds, and give CLAP the same audio transcript text; if the classical features then match or exceed the MLLM caption pipeline's HR@10 on MicroLens-100K, the claim that semantic captions are responsible for the gains would be contradicted.

Watch

Extended reading notes

Core claim

The central discovery: high-level semantics extracted by an off-the-shelf MLLM and fed as text embeddings outperform the low-level video, audio, and metadata features normally used in video recommendation. The pipeline prompts Qwen-VL zero-shot for video captions, uses Whisper plus Qwen-Audio for audio descriptions, encodes both with BGE-large, and feeds them into standard recommenders. In a two-tower model and in SASRec, the MLLM representations rank higher across HR@10/@30 and nDCG@10/@30. The largest gain is audio-to-text: two-tower HR@10 rises from 0.0253 to 0.0405, about 60% relative. Scaling the text encoder or MLLM backbone further does not help, suggesting the gains come from the sem

Load-bearing premise

The paper assumes that the performance gains come from the meaning captured in the generated captions, not from the captioning model simply seeing more of the video and audio or using a larger model than the baseline encoders.

Editorial extensions

If this is right

  • Audio-to-text conversion is the single largest win: roughly 60% relative HR@10 gain in the two-tower model, suggesting audio semantics are currently badly under-used.
  • MLLM video captions outperform creator-written titles by 4% to 18%, so the approach is a viable replacement or complement for unreliable metadata.
  • The gains appear in both an embedding-based two-tower model and the generative SASRec, so the improvement transfers across recommender architectures.
  • Because the MLLM is frozen and no finetuning is needed, the pipeline can be added to an existing system without retraining the content model.
  • Video captioning helps most for clips longer than 30 seconds, where fixed-frame video encoders miss most of the content.

Reading between the lines

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

  • A controlled comparison with matched frame counts, video coverage, and parameter counts would settle whether the gains are due to semantic content or to the MLLM's larger perceptual budget; the paper does not run this control.
  • Because captions are human-readable, the same pipeline could double as an interpretability layer, letting recommender systems explain to users or auditors why a clip was recommended.
  • The large audio gain suggests that transcribing and describing soundtracks is a particularly cheap win; platforms with clean audio metadata could test this first.
  • The diminishing returns from larger MLLMs hint that caption quality for recommendation saturates quickly; an accessible test is to compare multiple prompt templates and caption lengths to find that saturation point.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes a zero-finetuning framework that uses an off-the-shelf multimodal large language model (Qwen-VL) to generate natural-language captions for short videos, encodes these captions with a text encoder, and feeds the resulting embeddings into standard recommender architectures (two-towers and SASRec). On MicroLens-100K, the reported results show that MLLM-generated video captions improve HR@10 and nDCG@10 over conventional VideoMAE features, CLAP audio features, and title metadata, with the largest gains in the two-towers model. The authors argue that MLLM captions capture high-level semantics such as intent, humor, and world knowledge that are missed by low-level encoders, and that scaling the MLLM or text encoder does not bring further gains.

Significance. If the central claim holds, the paper offers a practical and inexpensive way to inject semantic understanding into video recommenders, using frozen MLLMs and standard text encoders. The approach is model-agnostic, reproducible via the released prompts and generated data, and the experimental setup uses a realistic time-based split. The paper is honest about not finetuning the MLLM and about using publicly available components. However, the contribution is currently weakened by a mismatch between the strong claims in the abstract/introduction and what is actually evaluated, and by a confounding in the comparison between MLLM video features and the VideoMAE baseline. The core direction is credible, but the evidence as presented does not yet establish that the gains come from semantic descriptions rather than from the MLLM's greater input coverage, larger capacity, or additional audio input.

major comments (4)
  1. [§5.1.2, §5.1.3, Table 1] The abstract and Section 1 claim that the framework was evaluated on 'five representative models', but Section 5.1.3 and Table 1 only report two models: two-towers and SASRec. This discrepancy is not merely a wording issue: the introduction also claims that 'multimodal captions, generated by analyzing both video and audio inputs jointly, yield the most robust representations', yet Table 1 contains no row for a concatenated audio+video MLLM caption. The empirical support for the central claims is therefore narrower than stated. The authors should either add the missing experiments or revise the claims to match the two evaluated models and the separately evaluated audio and video captions.
  2. [§5.1.2, Table 1] The comparison between MLLM video and VideoMAE is confounded. VideoMAE is a 307M model processing only 16 frames extracted from the first 30 seconds, while Qwen-VL is a much larger vision-language model that can process frames sampled throughout clips up to 4 minutes, and the pipeline also feeds audio transcriptions from Whisper and Qwen-Audio into the caption generation. The paper states the encoders have 'comparable parameter counts', but Qwen-VL is substantially larger than VideoMAE (and Table 2 explicitly mentions a 7B variant of Qwen-VL). No control condition is reported where VideoMAE sees more frames or the full clip, nor is there an audio+video conventional baseline. Consequently, the observed gains (e.g., HR@10 0.0393→0.0489) could be due to more frames, audio input, or model capacity, rather than to the semantic nature of the captions. This attribution problem should be address
  3. [Table 1] The two-towers MLLM video row reports nDCG@30 = 0.355, which is outside the [0,1] range for nDCG and inconsistent with the nDCG@10 value of 0.0264 in the same row. This appears to be a typo (likely 0.0355), but as printed it undermines the credibility of the results table. Please correct and check all other entries for similar decimal-point errors.
  4. [§5.2, Table 2] The paper claims that 'merely increasing model size offers diminishing returns', but Table 2 reports only two scaling experiments (larger text encoder and larger MLLM) without statistical significance tests. The absence of gains could also be due to the captions already saturating the downstream recommender capacity, or to the larger models not being applied to the same prompt as the baseline. The conclusion about diminishing returns is not strongly supported by the evidence presented.
minor comments (6)
  1. [§5.1.2] The text says encoders are selected with 'comparable parameter counts', but the numbers cited (CLAP 194M, VideoMAE 307M, BGE-large 335M) are not comparable to Qwen-VL, whose parameter count is not given. Please state the Qwen-VL model size explicitly and either remove or justify the comparability claim.
  2. [§5.1.2] It is not explicitly stated that the same text encoder (BGE-large) is used for MLLM-generated captions and for metadata. Since the text encoder is a component of the framework, this should be stated for reproducibility.
  3. [Table 1] No significance indicators (e.g., *, †) are shown despite the text saying all comparisons are based on a paired t-test. Please add significance markers or report confidence intervals.
  4. [§5.1.2, References] The paper refers to 'Qwen-VL' but the cited reference [25] is 'Qwen2-VL'. Please clarify the exact model version used (e.g., Qwen2-VL-2B) and its parameter count.
  5. [§5.2] The sentence describing the SASRec audio gain says 'a relative gain of 35%', which is approximately correct (0.0338→0.0454), but it would be clearer to show the exact computation or cite the table row.
  6. [§5.1.1] Please report how many videos remain after filtering to a maximum length of 4 minutes and downscaling, since this affects the comparability of the results with the original MicroLens-100K numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: MLLM captions come from frozen public models and recommender scores are computed on held-out interactions, so the reported gains are not true by construction.

full rationale

The paper's chain is a standard empirical comparison, not a derivation that re-imports its own output. Captions are generated by off-the-shelf frozen MLLMs (Qwen-VL, Whisper, Qwen-Audio), embedded with a fixed text encoder (BGE-large), and fed into standard recommenders (two-towers, SASRec) trained on the MicroLens-100K global time split and evaluated on held-out days. No parameter is fitted to the evaluation labels and then reported as a prediction; no normalization, scaling, or selection rule is defined in terms of the target HR/nDCG values; and the recommender hyperparameters are taken from prior work without tuning. There is no load-bearing self-citation: the reference list is external (VideoMAE, SASRec, Qwen-VL, MicroLens, etc.), and the only author-provided artifact is a released dataset/prompt link, which is not used as evidence for the empirical claims. The caveats noted by the reader—VideoMAE consumes only 16 frames from the first 30 seconds while the MLLM receives more frames plus audio, and the abstract says 'five representative models' while Table 1 reports two—are correctness/attribution concerns about confounds and reporting consistency, not circularity: the comparison could have favored the baselines, and nothing in the method forces the MLLM features to win. The nDCG@30 value 0.355 in Table 1 appears to be a typographical inconsistency, but it is not a circular step. Overall, the central claim is self-contained against external baselines and receives a score of 0.

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

No free parameters were fitted for the recommendation models; hyperparameters are inherited from prior work (Section 5.1.3). The main axioms are the representativeness of MicroLens-100K, the semantic quality of frozen MLLM captions, and fairness of the conventional encoders as baselines.

assumptions (3)
  • domain assumption MLLM captions preserve high-level semantics (intent, humor, world knowledge) that correlate with user engagement
    Core hypothesis stated in Section 3; not independently verified besides downstream ranking gains.
  • domain assumption MicroLens-100K interactions and the 1-day rolling time split represent production short-video recommendation
    Section 5.1.1 and 5.1.4; single dataset, no cross-dataset evidence.
  • domain assumption Baseline encoders VideoMAE (16 frames, first 30s), CLAP, and BGE titles are fair comparators for isolating the MLLM effect
    Section 5.1.2; no ablation controls MLLM capacity, frame count, or audio transcription information.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Describe What You See with Multimodal Large Language Models to Enhance Video Recommendations." pith.science (2026). https://pith.science/paper/BYHYKNPC

@misc{pith2026250809789,
  author       = {Pith},
  title        = {Pith review of: Describe What You See with Multimodal Large Language Models to Enhance Video Recommendations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BYHYKNPC}},
  note         = {Machine review of arXiv:2508.09789}
}
read the original abstract

Existing video recommender systems rely primarily on user-defined metadata or on low-level visual and acoustic signals extracted by specialised encoders. These low-level features describe what appears on the screen but miss deeper semantics such as intent, humour, and world knowledge that make clips resonate with viewers. For example, is a 30-second clip simply a singer on a rooftop, or an ironic parody filmed amid the fairy chimneys of Cappadocia, Turkey? Such distinctions are critical to personalised recommendations yet remain invisible to traditional encoding pipelines. In this paper, we introduce a simple, recommendation system-agnostic zero-finetuning framework that injects high-level semantics into the recommendation pipeline by prompting an off-the-shelf Multimodal Large Language Model (MLLM) to summarise each clip into a rich natural-language description (e.g. "a superhero parody with slapstick fights and orchestral stabs"), bridging the gap between raw content and user intent. We use MLLM output with a state-of-the-art text encoder and feed it into standard collaborative, content-based, and generative recommenders. On the MicroLens-100K dataset, which emulates user interactions with TikTok-style videos, our framework consistently surpasses conventional video, audio, and metadata features in five representative models. Our findings highlight the promise of leveraging MLLMs as on-the-fly knowledge extractors to build more intent-aware video recommenders.

Figures

Figures reproduced from arXiv: 2508.09789 by the authors.

Figure 1
Figure 1. A qualitative breakdown of a gameplay clip from a Naruto mobile fighting game (video ID 9183), demonstrating the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 14 canonical work pages

  1. [1]

    Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. 2016. Youtube-8m: A large-scale video classification benchmark.arXiv preprint arXiv:1609.08675 (2016)

    arXiv 2016

  2. [2]

    Kirolos Ataallah, Xiaoqian Shen, Eslam Abdelrahman, Essam Sleiman, Deyao Zhu, Jian Ding, and Mohamed Elhoseiny. 2024. Minigpt4-video: Advancing multimodal llms for video understanding with interleaved visual-textual tokens. arXiv preprint arXiv:2404.03413(2024)

    arXiv 2024

  3. [3]

    Yunfei Chu, Jin Xu, Xiaohuan Zhou, Qian Yang, Shiliang Zhang, Zhijie Yan, Chang Zhou, and Jingren Zhou. 2023. Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models.arXiv preprint arXiv:2311.07919(2023)

    arXiv 2023

  4. [4]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep neural networks for youtube recommendations. InProceedings of the 10th ACM conference on recommender systems. 191–198

    2016

  5. [5]

    Chaoyou Fu, Haojia Lin, Xiong Wang, Yi-Fan Zhang, Yunhang Shen, Xiaoyu Liu, Haoyu Cao, Zuwei Long, Heting Gao, Ke Li, et al. 2025. Vita-1.5: Towards gpt-4o level real-time vision and speech interaction.arXiv preprint arXiv:2501.01957 (2025)

    arXiv 2025

  6. [6]

    Junchen Fu, Xuri Ge, Xin Xin, Alexandros Karatzoglou, Ioannis Arapakis, Kaiwen Zheng, Yongxin Ni, and Joemon M Jose. 2024. Efficient and Effective Adaptation of Multimodal Foundation Models in Sequential Recommendation.arXiv preprint arXiv:2411.02992(2024)

    work page Pith review arXiv 2024

  7. [7]

    Xudong Gong, Qinlin Feng, Yuan Zhang, Jiangling Qin, Weijie Ding, Biao Li, Peng Jiang, and Kun Gai. 2022. Real-time Short Video Recommendation on Mobile Devices. InProceedings of the 31st ACM International Conference on Information & Knowledge Management. 3103–3112

    2022

  8. [8]

    Pan Gu, Haiyang Hu, and Guandong Xu. 2024. Modeling multi-behavior sequence via HyperGRU contrastive network for micro-video recommendation.Knowledge- Based Systems295 (2024), 111841

    work page 2024

Show all 42 references
  1. [9]

    Hao Jiang, Wenjie Wang, Yinwei Wei, Zan Gao, Yinglong Wang, and Liqiang Nie

  2. [10]

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

    2018

  3. [11]

    Joonseok Lee and Sami Abu-El-Haija. 2017. Large-scale content-only video recommendation. InProceedings of the IEEE International Conference on Computer Vision Workshops. 987–995

    2017

  4. [12]

    Chenyi Lei, Yong Liu, Lingzi Zhang, Guoxin Wang, Haihong Tang, Houqiang Li, and Chunyan Miao. 2021. Semi: A sequential multi-modal information transfer network for e-commerce micro-video recommendations. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & ...

    2021

  5. [13]

    Youhua Li, Hanwen Du, Yongxin Ni, Pengpeng Zhao, Qi Guo, Fajie Yuan, and Xi- aofang Zhou. 2024. Multi-modality is all you need for transferable recommender systems. In2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 5008–5021

    2024

  6. [14]

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2023. Improved Baselines with Visual Instruction Tuning

    2023

  7. [15]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual Instruc- tion Tuning. InNeurIPS

    2023

  8. [16]

    Zaiqiao Meng, Richard McCreadie, Craig Macdonald, and Iadh Ounis. 2020. Ex- ploring data splitting strategies for the evaluation of recommendation models. In Proceedings of the 14th acm conference on recommender systems. 681–686

    2020

  9. [17]

    Yongxin Ni, Yu Cheng, Xiangyan Liu, Junchen Fu, Youhua Li, Xiangnan He, Yongfeng Zhang, and Fajie Yuan. 2023. A content-driven micro-video recommen- dation dataset at scale.arXiv preprint arXiv:2309.15379(2023)

    2023 arXiv

  10. [18]

    Nikil Pancha, Andrew Zhai, Jure Leskovec, and Charles Rosenberg. 2022. Pinner- former: Sequence modeling for user representation at pinterest. InProceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining. 3702–3712

    2022

  11. [19]

    Peixuan Qi. 2024. Movie Visual and Speech Analysis through Multi-Modal LLM for Recommendation Systems.IEEE Access(2024)

    2024

  12. [20]

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2023. Robust speech recognition via large-scale weak supervision. InInternational conference on machine learning. PMLR, 28492–28518

    2023

  13. [21]

    Anima Singh, Trung Vu, Nikhil Mehta, Raghunandan Keshavan, Maheswaran Sathiamoorthy, Yilin Zheng, Lichan Hong, Lukasz Heldt, Li Wei, Devansh Tandon, et al. 2024. Better generalization with semantic ids: A case study in ranking for recommendations. InProceedings of the 18th ACM...

    2024

  14. [22]

    Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. 2022. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems35 (2022), 10078–10093

    2022

  15. [23]

    Elahe Vahdani and Yingli Tian. 2022. Deep learning-based action detection in untrimmed videos: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence45, 4 (2022), 4302–4320

    2022

  16. [24]

    Karan Vombatkere, Sepehr Mousavi, Savvas Zannettou, Franziska Roesner, and Krishna P Gummadi. 2024. Tiktok and the art of personalization: Investigating exploration and exploitation on social media feeds. InProceedings of the ACM Web Conference 2024. 3789–3797

    2024

  17. [25]

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. 2024. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191(2024)

    2024 arXiv

  18. [26]

    Yinwei Wei, Xiang Wang, Liqiang Nie, Xiangnan He, Richang Hong, and Tat-Seng Chua. 2019. MMGCN: Multi-modal graph convolution network for personalized recommendation of micro-video. InProceedings of the 27th ACM international conference on multimedia. 1437–1445

    2019

  19. [27]

    Chao-Yuan Wu, Yanghao Li, Karttikeya Mangalam, Haoqi Fan, Bo Xiong, Jiten- dra Malik, and Christoph Feichtenhofer. 2022. Memvit: Memory-augmented multiscale vision transformer for efficient long-term video recognition. InPro- ceedings of the IEEE/CVF Conference on Computer Vis...

    2022

  20. [28]

    Yusong Wu, Ke Chen, Tianyu Zhang, Yuchen Hui, Taylor Berg-Kirkpatrick, and Shlomo Dubnov. 2023. Large-scale contrastive language-audio pretraining with feature fusion and keyword-to-caption augmentation. InICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and...

    2023

  21. [29]

    Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. 2023. C-Pack: Packaged Resources To Advance General Chinese Embedding. arXiv:2309.07597 [cs.CL]

    2023 arXiv

  22. [30]

    Jin Xu, Zhifang Guo, Jinzheng He, Hangrui Hu, Ting He, Shuai Bai, Keqin Chen, Jialin Wang, Yang Fan, Kai Dang, et al. 2025. Qwen2. 5-omni technical report. arXiv preprint arXiv:2503.20215(2025)

    2025 arXiv

  23. [31]

    Yuyang Ye, Zhi Zheng, Yishan Shen, Tianshu Wang, Hengruo Zhang, Peijun Zhu, Runlong Yu, Kai Zhang, and Hui Xiong. 2025. Harnessing multimodal large language models for multimodal sequential recommendation. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 3...

    2025

  24. [33]

    Yisong Yu, Beihong Jin, Jiageng Song, Beibei Li, Yiyuan Zheng, and Wei Zhuo

  25. [34]

    Guanghu Yuan, Fajie Yuan, Yudong Li, Beibei Kong, Shujie Li, Lei Chen, Min Yang, Chenyun Yu, Bo Hu, Zang Li, et al. 2022. Tenrec: A Large-scale Multipurpose Benchmark Dataset for Recommender Systems.arXiv preprint arXiv:2210.10629 (2022)

    2022 arXiv

  26. [35]

    Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, et al. 2025. Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models.arXiv preprint arXiv:2506.05176(2025)

    2025 arXiv

  27. [36]

    Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jen- nifer J

    Long Zhao, Nitesh B. Gundavarapu, Liangzhe Yuan, Hao Zhou, Shen Yan, Jen- nifer J. Sun, Luke Friedman, Rui Qian, Tobias Weyand, Yue Zhao, Rachel Hornung, Florian Schroff, Ming-Hsuan Yang, David A. Ross, Huisheng Wang, Hartwig Adam, Mikhail Sirotenko, Ting Liu, and Boqing Gong....

    2024

  28. [37]

    Yu Zheng, Chen Gao, Jingtao Ding, Lingling Yi, Depeng Jin, Yong Li, and Meng Wang. 2022. Dvr: micro-video recommendation optimizing watch-time-gain under duration bias. InProceedings of the 30th ACM International Conference on Multimedia. 334–345

    2022

  29. [38]

    Ting Zhong, Jian Lang, Yifan Zhang, Zhangtao Cheng, Kunpeng Zhang, and Fan Zhou. 2024. Predicting Micro-video Popularity via Multi-modal Retrieval Augmentation. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2579–2583

    2024

  30. [39]

    Peilin Zhou, Chao Liu, Jing Ren, Xinfeng Zhou, Yueqi Xie, Meng Cao, Zhongtao Rao, You-Liang Huang, Dading Chong, Junling Liu, Jae Boum Kim, Shoujin Wang, Raymond Chi-Wing Wong, and Sunghun Kim. 2025. When Large Vision Language Models Meet Multimodal Sequential Recommendation: ...

    2025

  31. [40]

    Tinghui Zhu, Kai Zhang, Muhao Chen, and Yu Su. 2025. Is Extending Modality The Right Path Towards Omni-Modality?arXiv preprint arXiv:2506.01872(2025)

    2025 arXiv

  32. [2020]

    InProceedings of the 28th ACM International conference on Multimedia

    What aspect do you like: Multi-scale time-aware user interest modeling for micro-video recommendation. InProceedings of the 28th ACM International conference on Multimedia. 3487–3495

  33. [2022]

    In Joint European Conference on Machine Learning and Knowledge Discovery in Databases

    Improving micro-video recommendation by controlling position bias. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, 508–523

  34. [2023]

    In Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2022, Grenoble, France, September 19–23, 2022, Proceedings, Part I

    Improving Micro-video Recommendation by Controlling Position Bias. In Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2022, Grenoble, France, September 19–23, 2022, Proceedings, Part I. Springer, 508–523

    2022

Pith tools

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