Pith. sign in

REVIEW 5 major objections 8 minor 1 cited by

MIM: Multi-modal Content Interest Modeling Paradigm for User Behavior Modeling

T0 review · 5 major / 8 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read MIM aligns item images and titles with purchase interest, improving click-through prediction for existing user behavior models.

desk verdict Solid industrial multi-modal UBM paper; the recipe is new, the A/B claims are thinly documented, and the purchase-pair mechanism deserves a control before we buy the 'content interest' story. read the letter →

arxiv 2502.00321 v4 pith:4HHLKENM submitted 2025-02-01 cs.IR cs.AI

Bencheng Yan , Si Chen , Shichang Jia , Jianyu Liu , Yueran Liu , Chenghan Fu , Wanxian Guan , Hui Zhao
show 7 more authors
Xiang Zhang Kai Zhang Wenbo Su Pengjie Wang Jian Xu Bo Zheng Baolin Liu
This is my paper · ORCID
classification cs.IRcs.AI
keywords multi-modalrepresentationsCTRpredictionuserbehaviormodelingcontrastivelearningsupervisedfine-tuningrecommendationsystemse-commercesearchcold-start
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 claims that a three-stage paradigm called MIM can make ordinary ID-based user behavior models see item content, not just item IDs. The key move is to fine-tune pre-trained image and text encoders with contrastive supervision from user purchase behavior, so embeddings come to represent what users are interested in rather than merely what an item looks like. Those interest-aware embeddings are then plugged into existing CTR models through a content-interest-aware user behavior module that combines ID, content, and fused signals. If the claim holds, MIM is a general plug-in that improves a wide class of user behavior models, and it is cheap enough to run at platform scale. The paper reports offline AUC gains of 0.23 to 0.66 percentage points across baselines and online gains of 14.14% in CTR and 4.12% in RPM on a large e-commerce search system.

What carries the argument

The load-bearing object is the content-interest-aware supervised contrastive fine-tuning (C-SFT) stage, which converts user purchase behavior into a training signal for multimodal embeddings. It defines an interest pair as an image query and the item purchased after that query, encodes query image, item image, and item title with vision and language encoders, fuses image and title embeddings by an outer-product tensor fusion followed by an MLP, and optimizes an InfoNCE loss with two extensions: space-time negative sampling that enlarges the negative pool roughly two hundredfold from recent batches and other GPUs, and a multi-level loss that also aligns the query image with item image and item title separately. This stage is what shifts embeddings from describing what the content is to expressing what users are interested in. The companion efficiency mechanism is a representation center that precomputes item embeddings into a table, with real-time inference for new items and a decomposed parameter/computation machine layout.

What would settle it

A direct test would be to construct C-SFT pairs from image queries and purchased items while controlling for confounds, for example restricting to cases where the purchased item is a deliberate visual match rather than the cheapest or most popular result, and then check whether the AUC gains vanish when the pairs are replaced by random non-purchase image-item pairs matched on category and price. If the gains persist with confounded or random pairs, the improvement is not attributable to content-interest alignment.

Watch

Extended reading notes

Core claim

The central discovery is that the semantic gap between content embeddings and user interests can be closed by supervised contrastive fine-tuning on naturally occurring image-query-to-purchase pairs, and that the resulting embeddings transfer to CTR prediction as an additive module that works across many base models. The authors construct interest pairs from visual search logs, treat the purchased item as the positive sample, and train with an InfoNCE loss extended by space-time negative sampling and multi-level losses over image, text, and fused multimodal embeddings. Once embeddings are aligned with purchase interest, the CiUBM layer weights both ID and multimodal behavior histories by content similarity to the target item and concatenates ID, content, and fusion interest representations into the deep CTR head. The paper's claim is that this alignment, not simply adding multimodal features, is what produces consistent improvements across all tested baselines and datasets.

Load-bearing premise

The load-bearing premise is that a user's purchase of an item after an image search is a direct signal of interest in that item's image content; if the purchase is actually driven by price, seller reputation, position bias, or popularity, the contrastive fine-tuning learns those confounds instead.

Editorial extensions

If this is right

  • Any existing UBM, including average pooling, DIN, DIEN, BST, TWIN, and SIM, can be upgraded by appending the CiUBM module, with the paper reporting AUC gains from +0.0023 to +0.0066 across all tested models and datasets.
  • Cold-start and long-tail items benefit more: the paper reports larger AUC gains for the newest item sets, because content embeddings give new items a representable feature profile before interactions accumulate.
  • The representation center keeps deployment cost near the ID-only baseline: training computation rises only 0.59x and inference 0.387x over Base, instead of hundreds- or thousands-fold for real-time or end-to-end multimodal variants.
  • Content-sensitive categories such as clothing gain more than content-insensitive ones such as cars, consistent with the claim that the mechanism models visual and textual interest.
  • The paradigm transfers beyond search: applying it to a display advertising recommendation system yielded additional offline and online gains.

Reading between the lines

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

  • If purchase-as-interest holds, the same C-SFT recipe could be applied to other logged behavioral signals, such as add-to-cart, long dwell time, or repeat purchase, and the relative AUC gains across signal choices would reveal which behaviors are the cleanest content-interest labels.
  • The framework suggests a cheap way to audit multimodal recommender systems: compare embedding alignment quality on hand-labeled visual-match pairs against logged purchase pairs, and large divergence would indicate confounded training pairs.
  • Because the representation center decouples embedding extraction from CTR training, the C-SFT embeddings could be refreshed on a cadence independent of model retraining, raising a testable question about how quickly interest-aligned embeddings decay for fashion and seasonal categories.
  • The roughly two-hundredfold expansion of negatives through space-time sampling implies that very large negative pools may be a generic lever for other contrastive user-interest alignment tasks, though the paper does not test that transfer.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 8 minor

Summary. The paper proposes MIM, a three-stage paradigm for CTR prediction in e-commerce search. Stage 1 pre-trains vision and language foundation models on domain data (DDA) and aligns modalities (DMA). Stage 2, C-SFT, uses supervised contrastive fine-tuning on user interest pairs defined as <image query, item purchased after that query>, with multi-level InfoNCE losses over image, text, and fused embeddings. Stage 3, CiUBM, combines ID-based user behavior representations with multi-modal content interest modules and a fusion module, and a representation center precomputes embeddings for efficiency. Offline experiments on Taobao data compare six ID-based UBMs with and without MIM across eight categories, reporting AUC gains of +0.0023 to +0.0066; online A/B tests are reported with cumulative gains of +14.14% CTR and +4.12% RPM, and a code/dataset link is provided.

Significance. If the empirical claims hold, MIM is a practically useful plug-in for injecting multi-modal content signals into ID-based user behavior models, and the efficiency decomposition makes it deployable at industrial scale. The paper has several strengths: consistent AUC improvements across all six backbones and all eight sub-datasets, a clear three-stage decomposable design, a cold-start analysis showing larger gains on newest items, and a public release of code and data. The reported efficiency numbers are particularly valuable: adding MIM increases training GFLOPs by only 0.59x and inference GFLOPs by 0.387x relative to the ID-only base, thanks to the representation center. However, the central mechanism—that purchase behavior is a clean signal of content interest—is not adequately isolated from confounds, and the statistical reliability of both offline and online gains is not documented. The paper also does not compare against existing multi-modal or content-based UBM baselines, so the scope of the 'universal paradigm' claim is narrower than stated.

major comments (5)
  1. [§3.3, Eq. (5)] Equation (5) is written as L = -log( exp(sim(q, positive)) / sum_{j=0}^K exp(sim(q, negative_j)) ). In this form the denominator contains only negatives and omits the positive term, so it is not the standard InfoNCE objective. Standard InfoNCE would include exp(sim(q, positive)) inside the denominator as well. If the actual implementation uses the standard form, the equation is a typographical error; if the implementation matches the text, the loss is a different objective whose properties should be justified. Since Eq. (5) is the core of C-SFT, please correct it and state precisely which objective was optimized.
  2. [§3.3 (Q1) and §4.3.2, Table 4] The load-bearing assumption is that the positive pair <image query q, item i purchased after q> is a direct expression of content interest. This is not tested against the main confounds: display position, price, seller reputation, and item popularity, all of which can drive purchases and are correlated with observable content. Table 4 compares purchase, click, and category signals, but the category version is a cross-entropy classification signal, not a random-pair no-interest contrastive control, and none of the variants controls for popularity/price/position. Please add a random-pair control (e.g., matching each query with a random item from the same exposure distribution) and an analysis stratified by price or popularity to show that the C-SFT gains survive after removing these confounds. Without such a control, the reported AUC gains cannot be attributed to 'content-interest alignment' as claimed.
  3. [§4.1.3 and Table 2] The text says all experiments were run multiple times with different random seeds and averages are reported, but no standard deviations, confidence intervals, or significance tests appear anywhere. Because the headline offline gains are only +0.0023 to +0.0066 in AUC, they may be within run-to-run variation. Please report the variance across seeds and a paired significance test (or bootstrap) for at least the main Table 2 comparisons, or provide a reason why variance is negligible at this data scale.
  4. [§4.7] The online A/B report omits the experimental protocol: traffic split, number of users/impressions, duration, metric definitions, and statistical significance levels. It is also unclear how the cumulative '+14.14% CTR' relates to Table 7's 'All +5%' in a single release; state explicitly whether the headline number is a sum of three successive releases and give per-release protocol details. Without this information, the online claim is not independently checkable and should be described as a company-internal observation rather than a validated experimental result.
  5. [§4.2 and §2.2] All empirical comparisons are against ID-based UBMs (Avg Pooling, DIN, DIEN, BST, TWIN, SIM) with and without MIM. The paper does not compare MIM against any existing multi-modal or content-based UBM method (e.g., MISSRec or other modality-based recommender models cited in Section 2.2). This makes the claim that MIM is a 'universal and effective multi-modal UBM paradigm' broader than the evidence supports. Please add at least one content-based baseline that uses the same pre-trained FoM features (e.g., direct concatenation of MM embeddings into the deepCTR head) to position the contribution against current multi-modal approaches.
minor comments (8)
  1. [§3.3 (Q1)] Please specify the temporal and session definition of 'purchased after that query': is the purchase in the same search session, the same day, or a later session, and is the item from the corresponding search result page?
  2. [§3.3 (Extension 1)] The claim of 'nearly 200x more negative samples' should be derived from batch size N, number of GPUs P, and k; currently the reader cannot verify this factor from the text.
  3. [Table 1] Entries such as '60,75+ M' and '3,66+ M' are likely formatting errors for '60.75+ M' and '3.66+ M'; please correct the table.
  4. [§4.1.3] The hyperparameters alpha=0.5, beta=0.5, and k=10 are said to be grid-searched, but the grid range and sensitivity results are not reported; a small sensitivity table would help readers judge robustness.
  5. [Table 6] The Delta columns are difficult to read and contain at least one formatting error ('+0.38.7'); please clarify the reference for the relative difference and fix the numbers.
  6. [§4.7] The sentence 'we have continually achieved three releases' contains a typo ('achieved' is intended) and should be rephrased; also specify which release underlies Table 7.
  7. [§4.1.3 and References] The visual backbone EVA-2 is cited as [31] in Section 4.1.3, but reference [31] is ConvNeXt V2; the correct citation appears to be [12]. Please audit all reference-citation pairings.
  8. [Abstract / Code Release] The code and dataset are released only through a non-archival cloud drive link; for reproducibility, provide a persistent repository (e.g., GitHub or Zenodo) with versioned code and a dataset description.

Circularity Check

0 steps flagged · score 2.0 of 10

No substantive circularity: MIM is fine-tuned on purchase pairs but evaluated on click prediction, so the reported gains are not forced by construction.

full rationale

The claimed derivation chain is: (i) pretrain foundation models on domain images/titles (DDA+DMA); (ii) fine-tune with InfoNCE on <image query, purchased item> pairs (Eq. 5-6); (iii) feed the resulting MM embeddings into CiUBM alongside ID embeddings (Eq. 7-9); (iv) measure CTR AUC offline and CTR/RPM online. None of these steps reduces to its own input by definition. The C-SFT supervision signal is purchase behavior, while the downstream evaluation is click prediction; these are distinct labels, and Table 4 shows that the click-supervised variants (v2/v4) are actually worse than the purchase-supervised variants (v3/v5), so the reported gains are not a self-fulfilling consequence of training and testing on the same signal. The ablations vary the supervision signal (category/click/purchase) and loss (CE/contrastive), and all are evaluated on held-out CTR data, making the comparison empirical rather than definitional. The only author self-citation is [35] (APG) in a routine first-sentence citation for CTR prediction; it is not load-bearing, and no 'uniqueness theorem' or prior-work ansatz is imported to force the method's form. The representation-center efficiency numbers are engineering measurements. The weakest point, that purchase may encode price/position/popularity confounds rather than content interest, is a construct-validity or correctness risk, not circularity: it does not make Eq. 5 equal to Eq. 9's output or make the AUC gain a fitted parameter. Score 2 reflects the minor non-load-bearing self-citation and no substantive circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper introduces no invented entities, but its central claim relies on standard contrastive-learning assumptions, an unverified equation of purchase with content interest, and grid-selected hyperparameters. The cached embedding design further assumes that item content distributions are stable enough for offline precomputed vectors to serve online.

free parameters (3)
  • alpha (loss weight for text alignment) = 0.5
    Grid-searched in Section 4.1.3; balances L_q2txt in Eq. 6.
  • beta (loss weight for image alignment) = 0.5
    Grid-searched in Section 4.1.3; balances L_q2img in Eq. 6.
  • k (number of past batches for negative sampling) = 10
    Grid-searched in Section 4.1.3; controls ST-NSG negative set size.
assumptions (5)
  • standard math InfoNCE contrastive loss is a valid and sufficient objective for aligning positive pairs and separating negatives.
    Adopted from CLIP [25]; the paper gives no derivation or proof but relies on the standard property.
  • domain assumption Purchase actions in visual search reliably indicate user interest in item content.
    Section 3.3 Q1 selects purchase over click because it is 'clearer and more reliable'; if false, the C-SFT alignment is confounded.
  • domain assumption ID embeddings capture collaborative filtering signals that are complementary to content embeddings.
    Section 3.4 ID Interest Module assumes the existing UBM machinery remains useful when combined with MM embeddings.
  • domain assumption Precomputed multi-modal embeddings cached in the representation center remain valid for new items and for online serving.
    Section 3.5 replaces real-time foundation-model inference with a cached embedding table, requiring distributional stability over time.
  • domain assumption AUC improvements on offline Taobao data translate into the reported online CTR and RPM gains.
    Section 4.2 asserts 0.1pt AUC can yield several percentage points of online CTR; the online A/B evidence is company-internal and not independently verifiable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MIM: Multi-modal Content Interest Modeling Paradigm for User Behavior Modeling." pith.science (2026). https://pith.science/paper/4HHLKENM

@misc{pith2026250200321,
  author       = {Pith},
  title        = {Pith review of: MIM: Multi-modal Content Interest Modeling Paradigm for User Behavior Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4HHLKENM}},
  note         = {Machine review of arXiv:2502.00321}
}
read the original abstract

Click-Through Rate (CTR) prediction is a crucial task in recommendation systems, online searches, and advertising platforms, where accurately capturing users' real interests in content is essential for performance. However, existing methods heavily rely on ID embeddings, which fail to reflect users' true preferences for content such as images and titles. This limitation becomes particularly evident in cold-start and long-tail scenarios, where traditional approaches struggle to deliver effective results. To address these challenges, we propose a novel Multi-modal Content Interest Modeling paradigm (MIM), which consists of three key stages: Pre-training, Content-Interest-Aware Supervised Fine-Tuning (C-SFT), and Content-Interest-Aware UBM (CiUBM). The pre-training stage adapts foundational models to domain-specific data, enabling the extraction of high-quality multi-modal embeddings. The C-SFT stage bridges the semantic gap between content and user interests by leveraging user behavior signals to guide the alignment of embeddings with user preferences. Finally, the CiUBM stage integrates multi-modal embeddings and ID-based collaborative filtering signals into a unified framework. Comprehensive offline experiments and online A/B tests conducted on the Taobao, one of the world's largest e-commerce platforms, demonstrated the effectiveness and efficiency of MIM method. The method has been successfully deployed online, achieving a significant increase of +14.14% in CTR and +4.12% in RPM, showcasing its industrial applicability and substantial impact on platform performance. To promote further research, we have publicly released the code and dataset at https://pan.quark.cn/s/8fc8ec3e74f3.

Figures

Figures reproduced from arXiv: 2502.00321 by the authors.

Figure 1
Figure 1. An example of ID interest and content interest [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The framework of MIM. There are a total of three stages, including Pre-training, C-SFT, and CiUBM. Besides, a [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. The framework of representation center. obtain MM embedding from the embedding table without any extra time and GPU memory cost. Real-time Inference Module (RIM). The key problem of main￾taining the MM embedding table is that it cannot obtain the MM features of new items in time. However, in practice, it is a very common situation where new items are shown all the time. Thus, RIM is designed to address this problem.… view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Evaluation of the impact of different FoMs including [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: AUC gains by different item sets [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. MOON2.0: Dynamic Modality-balanced Multimodal Representation Learning for E-commerce Product Understanding

    cs.CV 2025-11 conditional novelty 5.0 of 10

    MOON2.0 combines modality-routed experts, intra-product image-text alignment, MLLM-generated data augmentation, and dynamic sample filtering to reach state-of-the-art zero-shot e-commerce product understanding.

Reference graph

Works this paper leans on

46 extracted references · 20 canonical work pages · cited by 1 Pith paper

  1. [1]

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al

  2. [2]

    Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. 2021. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254 (2021)

    arXiv 2021

  3. [3]

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

    2023

  4. [4]

    Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. 2020. Generative pretraining from pixels. In International conference on machine learning . PMLR, 1691–1703

    work page 2020

  5. [5]

    Qiwei Chen, Huan Zhao, Wei Li, Pipei Huang, and Wenwu Ou. 2019. Behavior sequence transformer for e-commerce recommendation in alibaba. InProceedings of the 1st international workshop on deep learning practice for high-dimensional sparse data. 1–4

    2019

  6. [6]

    Gonzalez, Ion Stoica, and Eric P

    Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90%* ChatGPT Quality. https://lmsys.org/blog/2023-03-30-vicuna/

    2023

  7. [7]

    Le, and Christopher D

    Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. 2020. ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020 . OpenReview.net. https://openreview.net/ forum?id=r1xMH1BtvB

    work page 2020

  8. [8]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep Neural Networks for YouTube Recommendations. Proceedings of the 10th ACM Conference on Recommender Systems (Sept. 2016), 191–198. doi:10.1145/2959100.2959190

    work page arXiv 2016

Show all 46 references
  1. [9]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805 (2018)

    2018 arXiv

  2. [10]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. doi:10.48550/arXiv.1810.04805 arXiv:1810.04805 [cs]

    2019 doi arXiv

  3. [11]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, G. Heigold, S. Gelly, Jakob Uszkoreit, and N. Houlsby. 2020. An Image Is Worth 16x16 Words: Transformers for Image Recognition at Scale. ArX...

    2020

  4. [12]

    Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. 2023. EVA-02: A Visual Representation for Neon Genesis. arXiv:2303.11331 [cs.CV]

    2023 arXiv

  5. [13]

    Tom Fawcett. 2006. An introduction to ROC analysis. Pattern recognition letters 27, 8 (2006), 861–874

    2006

  6. [14]

    Tong He, Zhi Zhang, Hang Zhang, Zhongyue Zhang, Junyuan Xie, and Mu Li

  7. [15]

    Qidong Liu, Jiaxi Hu, Yutian Xiao, Jingtong Gao, and Xiang Zhao. 2023. Multi- modal Recommender Systems: A Survey. ArXiv (2023)

    2023

  8. [16]

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin transformer: Hierarchical vision transformer us- ing shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision. 10012–10022

    2021

  9. [17]

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. 2022. A ConvNet for the 2020s. doi:10.48550/arXiv.2201.03545 arXiv:2201.03545 [cs]

    2022 doi arXiv

  10. [18]

    Xianzhong Long, Han Du, and Yun Li. 2024. Two momentum contrast in triplet for unsupervised visual representation learning. Multim. Tools Appl. 83, 4 (2024), 10467–10480. doi:10.1007/S11042-023-15998-3

    2024 doi

  11. [19]

    Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. 2019. Vilbert: Pretrain- ing task-agnostic visiolinguistic representations for vision-and-language tasks. (2019)

    2019

  12. [20]

    Corrado, and Jeffrey Dean

    Tomás Mikolov, Ilya Sutskever, Kai Chen, Gregory S. Corrado, and Jeffrey Dean. 2013. Distributed Representations of Words and Phrases and their Compositionality. In Advances in Neural Information Processing Systems 26: 27th Annual Conference on Neural Information Processing Sy...

    2013

  13. [21]

    Kaixiang Mo, Bo Liu, Lei Xiao, Yong Li, and Jie Jiang. 2015. Image feature learning for cold start problem in display advertising. In Twenty-Fourth International Joint Conference on Artificial Intelligence

    2015

  14. [22]

    OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, and Lama Ahmad... 2024. GPT-4 Technical Report. arXiv:2303.08774 [cs.CL] https://arxiv.org/abs/2303. 08774

    2024 arXiv

  15. [23]

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

    2020

  16. [24]

    Jiezhong Qiu, Qibin Chen, Yuxiao Dong, Jing Zhang, Hongxia Yang, Ming Ding, Kuansan Wang, and Jie Tang. 2020. GCC: Graph Contrastive Coding for Graph Neural Network Pre-Training. In KDD ’20: The 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, ...

    2020

  17. [25]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

    2021

  18. [26]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners. OpenAI blog 1, 8 (2019), 9

    2019

  19. [27]

    Hao Tan and Mohit Bansal. 2019. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490 (2019)

    2019 arXiv

  20. [28]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guil- laume Lample. 2023. LLaMA: Open and Efficient Foundation ...

    2023 arXiv

  21. [29]

    Jinpeng Wang, Ziyun Zeng, Yunxiao Wang, Yuting Wang, Xingyu Lu, Tianxiang Li, Jun Yuan, Rui Zhang, Hai-Tao Zheng, and Shu-Tao Xia. 2023. MISSRec: Pre- training and Transferring Multi-modal Interest-aware Sequence Representation for Recommendation. In Proceedings of the 31st AC...

    2023

  22. [30]

    Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al

  23. [31]

    Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, and Saining Xie. 2023. ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders. arXiv:2301.00808 [cs.CV]

    2023 arXiv

  24. [32]

    Chuhan Wu, Fangzhao Wu, Tao Qi, and Yongfeng Huang. 2021. Empowering news recommendation with pre-trained language models. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Infor- mation Retrieval. 1652–1656

    2021

  25. [33]

    arXiv preprint arXiv:2208.10442 (2022)

    Image as a foreign language: Beit pretraining for all vision and vision- language tasks. arXiv preprint arXiv:2208.10442 (2022)

    2022 arXiv

  26. [34]

    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

  27. [35]

    Bencheng Yan, Pengjie Wang, Kai Zhang, Feng Li, Hongbo Deng, Jian Xu, and Bo Zheng. 2022. Apg: Adaptive parameter generation network for click-through rate prediction. Advances in Neural Information Processing Systems 35 (2022), 24740–24752

    2022

  28. [36]

    Yu, and Dahua Lin

    Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. 2018. Unsupervised Feature Learning via Non-Parametric Instance Discrimination. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018. Computer Vision Fou...

    2018

  29. [37]

    Fajie Yuan, Xiangnan He, Alexandros Karatzoglou, and Liguang Zhang. 2020. Parameter-Efficient Transfer from Sequential Behaviors for User Modeling and Recommendation. arXiv:2001.04253 [cs]

    2020 arXiv

  30. [38]

    Fajie Yuan, Guoxiao Zhang, Alexandros Karatzoglou, Joemon Jose, Beibei Kong, and Yudong Li. 2021. One Person, One Model, One World: Learning Continual User Representation without Forgetting. Proceedings of the 44th International ACM SIGIR Conference on Research and Development...

    2021

  31. [39]

    Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. Advances in neural information processing systems 32 (2019)

    2019

  32. [40]

    Amir Zadeh, Minghai Chen, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. 2017. Tensor Fusion Network for Multimodal Sentiment Analysis. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing. 1103–1114

    2017

  33. [41]

    Weinan Zhang, Jiarui Qin, Wei Guo, Ruiming Tang, and Xiuqiang He. 2021. Deep learning for click-through rate estimation.arXiv preprint arXiv:2104.10584 (2021)

    2021 arXiv

  34. [42]

    Zheng Yuan, Fajie Yuan, Yu Song, Youhua Li, Junchen Fu, Fei Yang, Yunzhu Pan, and Yongxin Ni. 2023. Where to Go Next for Recommender Systems? ID- vs. Modality-based Recommender Models Revisited. Proceedings of the 46th International ACM SIGIR Conference on Research and Develop...

    2023

  35. [43]

    Guorui Zhou, Xiaoqiang Zhu, Chenru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep interest network for click-through rate prediction. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining ...

    2018

  36. [45]

    Guorui Zhou, Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, and Kun Gai. 2019. Deep interest evolution network for click-through rate Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Yan et al. prediction. In Proceedings of the AAAI conference on artif...

    2019

  37. [2019]

    In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Bag of tricks for image classification with convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 558–567

  38. [2022]

    Flamingo: a visual language model for few-shot learning. (2022)

    2022

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.