Pith. sign in

REVIEW 3 major objections 5 minor 52 references

Popularity prediction improves when retriever ranks posts by social attributes as well as content.

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 · deepseek-v4-flash

2026-08-01 13:06 UTC pith:7XNT7NEZ

load-bearing objection The relation-guided transformer idea is a real extension and the ablations support it, but the headline results are not credible until the retrieval database split is specified—Eq. 18 feeds true labels into the predictor—and the abstract overclaims SOTA on ICIP. the 3 major comments →

arxiv 2607.19200 v1 pith:7XNT7NEZ submitted 2026-07-21 cs.MM

Enhancing Relation Modeling with Social Attributes for Social Media Popularity Prediction

Bolun Zheng , Yuhao Luo , Wei Zhu , Ning Xu , An-An Liu , Lingyu Zhu , Canjin Wang This is my paper
classification cs.MM
keywords social media popularity predictionretrieval-augmented generationsocial attributesrelation modelingmultimodal fusiontransformerUGCpopularity forecasting
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.

The paper proposes RE-Rag, a retrieval-augmented framework for social media popularity prediction. It argues that semantic similarity alone is a poor guide for selecting helpful historical posts, because similar content can spread through very different social contexts. The framework therefore retrieves posts using both multimodal content similarity and a rarity-weighted compensation based on social attributes such as author, category, and language. It then feeds the retrieved posts through transformers whose attention is modulated by explicit relation maps encoding whether pairs of posts share attribute values with the target. On three public benchmarks, RE-Rag reports the best accuracy, including relative improvements of 29.61%, 29.43%, and 3.80% in MSE, MAE, and SRC over the second-best method on the SMPD dataset, along with much faster retrieval than prior retrieval-augmented systems.

Core claim

The paper's central claim is that social media popularity is driven by two independent factors: content quality and the platform's distribution mechanism, with social attributes serving as observable proxies for that distribution mechanism. On this basis, RE-Rag treats retrieval selection as needing both semantic alignment and attribute alignment, and treats the predictor as needing to exploit relative attribute relations among the target and retrieved posts. The Semantic-Attribute Retriever multiplies the multimodal semantic similarity by a compensation weight derived from the global and local rarity of each social attribute, and the Relation-Guided Transformer adjusts attention by multiply

What carries the argument

The core mechanism is a pair of modules. First, the Semantic-Attribute Retriever (SAR) scores candidate posts by the product of a cosine multimodal semantic similarity and a compensation weight based on the rarity of each social attribute, combining a global IDF-like rarity and a local rarity within each attribute to upweight rare attribute values. Second, the Relation-Guided Predictor contains two Relation-Guided Transformers (s-RGT and c-RGT) that modify the standard attention computation: attribute-wise relation maps assign a value of 0, 1, or 2 to each pair of posts depending on whether their attribute values match each other and the target, then an embedding and MLP turn the maps into a

Load-bearing premise

The claim rests on the requirement that the retrieval database contains only training-set instances and never the query target or its popularity label; if the target's own popularity can appear among the retrieved features, the reported gains may reflect information leakage rather than genuine predictive skill.

What would settle it

Run the released code with a retrieval database explicitly built from the training split only, confirm that test instances are excluded, then ablate the p_i (popularity) and similarity features in Eq. 18: if removing the popularity of retrieved posts collapses the reported improvements, the method has been relying on leaked labels. A simpler sanity check is to compare performance when the database is restricted to training posts versus when it may contain test posts.

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

If this is right

  • If RE-Rag is correct, retrieval-augmented popularity predictors should treat social attributes as proxies for the distribution mechanism rather than as auxiliary content cues.
  • Retrieval quality improves when candidate posts are selected for popularity-distribution consistency, not just semantic similarity, as evidenced by smaller popularity gaps in the case study.
  • The relation-guided transformer shows that explicitly encoding relative relations among retrieved instances can outperform standard transformer attention for sequence-based prediction features.
  • The large performance gap on the SMPD dataset suggests attribute-aware retrieval yields the biggest benefits on large, diverse, author-rich datasets.
  • The reported retrieval time of 72 seconds versus hours for baselines indicates that the attribute-compensation step is computationally cheap while improving accuracy.

Where Pith is reading between the lines

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

  • The dual-factor framing suggests a testable decomposition: one could predict a content-quality score and a distribution-exposure score separately, then combine them, and compare against RE-Rag's interaction-based approach.
  • A key risk not addressed in the paper is label leakage: if the retrieval database contains the test post itself or its popularity value, the p_i and similarity features fed into the predictor (Eq. 18) could leak the answer. A reader should verify that the database is restricted to training-set instances and that gains persist when retrieved popularity values are withheld.
  • The rarity weight treats all attribute mismatches as equal, but real distribution effects are likely graded; learning attribute-specific importance weights would be a natural extension.
  • The relation map discretizes attribute similarity into three levels; using continuous or semantic attribute distances could capture finer propagation patterns, though this is beyond the paper's scope.

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 / 5 minor

Summary. The paper proposes RE-Rag, a retrieval-augmented framework for social-media popularity prediction. It introduces a Semantic-Attribute Retriever that scores candidate posts by a weighted combination of image/text similarity and an attribute-rarity compensation term, and a Relation-Guided Predictor that uses a relation map over social attributes to modulate self- and cross-attention in a transformer. Experiments on ICIP, SMPD, and SMTPD report state-of-the-art MSE/MAE/SRC on SMPD and SMTPD, with relative gains of 29.61%, 29.43%, and 3.80% over the second-best method on SMPD. Ablations and retrieval-quality analyses are provided, and code/data are released.

Significance. If the reported gains are genuine, the paper makes a useful contribution by showing that explicit modeling of relative social-attribute relations can improve both the retrieval stage and the prediction stage of retrieval-augmented popularity prediction. The framework is simple, well-motivated, and reproducible in principle (code and data are promised). The ablation study is reasonably thorough and supports the contribution of the relation-guided transformer. However, the central empirical claim hinges on an unspecified detail of the retrieval database construction, which currently leaves a plausible label-leakage path that could fully account for the headline improvements. Until that is resolved, the state-of-the-art claim is not credible.

major comments (3)
  1. [§3.2.1, §4.1.1, Eq. (18)] The retrieval database composition is not specified. Section 3.2.1 states each entry is <content, attribute, label>, where 'label' is the popularity; Section 4.1.1 gives an 8:1:1 random split for training/validation/test but never states that the retrieval database is built from the training split only. Since Eq. (18) concatenates the true popularity p_i of every retrieved item into the predictor input, any test or validation instance in the candidate pool leaks its ground-truth label into the feature vector at evaluation time. If the query post itself can appear in its own candidate pool, the problem becomes near-trivial. The paper must state explicitly that the retrieval database contains only training instances, and that test queries are excluded. In addition, because the split is described as '8:1:1' without a chronological ordering, 'historical' popularity of retrieved posts is not
  2. [§4.2, Table 2] The claim that RE-Rag 'consistently outperforms state-of-the-art methods' (Abstract) and 'outperforms on the vast majority of metrics' is not supported by the paper's own Table 2 on the ICIP dataset. There, SKAPP achieves better MSE (1.1075 vs. 1.1360) and better SRC (0.6975 vs. 0.6740), and RE-Rag is second-best. The reader is told that this is because ICIP has 'weaker distribution-related attributes,' but no quantitative support is given. Moreover, although the paper states that each model is run five times, no error bars, standard deviations, or significance tests are reported anywhere. The reported differences—especially the 3.80% SRC gain on SMPD and the SMTPD gains—may be within run-to-run variability, which the current presentation does not rule out. Please report variances and perform significance tests (or at least error bars) for the main comparisons, and temper the 'consistent
  3. [§4.4, Fig. 3] The parameter-sensitivity analysis is hard to evaluate as presented. The figure contains garbled text with /uni... sequences rather than readable axis labels, and the surrounding text references 'Fig. 3 (b)' but the figure is not legible. The claim that N_retrieval=40 and alpha=0.3 are chosen to 'balance' performance would be more convincing if the curves were readable and if the chosen values were clearly marked. This is not a correctness issue but does impair verification of a stated hyperparameter choice.
minor comments (5)
  1. [§4.2 text] The sentence 'SKAPP achieves competitive performance on the smaller ICIP dataset...' is an interpretation, not a result. Consider reporting a direct comparison or analysis of attribute statistics across datasets to support it.
  2. [General] There are typesetting artifacts in Figure 3 and in Table 4/Figure 5 captions (e.g., /uni00000015... sequences). These should be cleaned before publication.
  3. [References] Reference [6] is cited as 'In-context Prompt-augmented Micro-video Popularity Prediction' but the citation style in the text uses [6] without an author list; please ensure the reference list is complete and consistent.
  4. [§3.3.2] Eq. (15) defines the relation map with three values {0,1,2}. This is an ad-hoc ordinal encoding and its adequacy is not discussed. It is acceptable empirically if ablations show benefit, but the paper should acknowledge that the mapping is a modeling choice rather than a derived result.
  5. [§4.1.2] The statement 'Each model is run five times, and all reported results are averaged over these runs' is good, but the paper never reports the standard deviation. Please include this information in tables or in a footnote.

Circularity Check

0 steps flagged

No circularity: RE-Rag is a standard supervised retrieval-augmented pipeline; the retrieval-database split ambiguity is an experimental-risk concern, not an in-paper circular reduction.

full rationale

The paper's derivation chain is not circular. The Semantic-Attribute Retriever (SAR) computes cosine similarities (Eqs. 2-4) and attribute compensation weights (Eqs. 5-11) from content features and social attribute values, neither of which includes the target popularity. The Relation-Guided Predictor (RGP) concatenates each retrieved instance's semantic feature with that instance's popularity p_i and compensated matching score (Eq. 18), then uses attention and MLP to predict the target's popularity (Eq. 20). The target's own label nowhere appears in the construction of its input features; p_i is the popularity of other, retrieved instances, which is exactly the auxiliary historical information retrieval-augmented methods are designed to exploit. The L2 loss supervises the prediction using the target label, but no equation defines the prediction as equivalent to a fitted quantity or to the label itself. The potential concern that the retrieval database may contain test instances (so that p_i could be the target's own ground-truth popularity) is a legitimate experimental-validity risk, but the manuscript does not state that the database is built from the test split, and the 8:1:1 split description does not force that interpretation. This is a data-split ambiguity to check in the released code, not a circular derivation demonstrated within the paper. Self-citations appear for the SMTPD benchmark [49], but that is a dataset contribution, not a load-bearing theoretical premise, and no uniqueness theorem or prior-work ansatz is invoked to make the model's choices forced. Thus the central claim, if the evaluation setup is correct, has independent empirical content.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The only free parameters are hyper-parameters tuned on validation. No new physical or conceptual entities are posited; the relation map and RGTs are internal architectural components. The main unstated assumption is the cleanliness of the retrieval database split.

free parameters (2)
  • alpha (retrieval modality weight) = 0.3
    Hyper-parameter balancing image and text similarity in Eq. (4); set to 0.3 after sensitivity analysis in Section 4.4.2.
  • N_retrieval (number of retrieved instances) = 40
    Set to 40 in Section 4.4.1 based on sensitivity; affects retrieval and prediction.
axioms (5)
  • domain assumption Social attributes (author, category, language) are valid proxies for the platform's content distribution mechanism.
    Stated in the Introduction; used to justify attribute compensation and relation maps. No direct validation that attributes capture exposure pathways.
  • domain assumption The retrieval database used at test time contains only historical posts and is disjoint from the query/test instances.
    Never stated in Section 3.2.1; if violated, the use of retrieved popularity p_i and similarity scores would leak the target label.
  • domain assumption Cosine similarity of pre-trained image/text features and IDF-based rarity are sufficient operationalizations of content and attribute relevance.
    Used in Eqs. (2)-(8); chosen heuristically.
  • domain assumption The popularity score p=log2(v/(d+1)) appropriately normalizes for time.
    Adopted from prior work [49] to align different post ages.
  • ad hoc to paper Relation map values (0,1,2) adequately represent the relative attribute relationships for attention modulation.
    Eq. (15) encodes equality with target as 2, equality among retrieved as 1, and inequality as 0; this quantization is a design choice without theoretical justification.

pith-pipeline@v1.3.0-alltime-deepseek · 16959 in / 10481 out tokens · 105316 ms · 2026-08-01T13:06:46.204758+00:00 · methodology

0 comments
read the original abstract

Recent studies highlight the critical role of retrieval-augmented mechanisms in social media popularity prediction (SMPP). Although such frameworks have improved SMPP performance by leveraging historical posts, existing methods still suffer from the low retrieval accuracy due to the oversight of relative relationships among UGC instances. To address this limitation, we propose a novel Relation-Enhanced Retrieval-Augmented framework (RE-Rag) that models UGC similarity as a continuous relation jointly driven by semantic content and social attributes. Specifically, RE-Rag employs a Semantic-Attribute Retriever (SAR) to obtain instances aligned in both semantic and social-attribute distributions. Subsequently, we design a Relation-Guided Predictor (RGP): first, cross-attention encodes multimodal features of retrieved instances; then, a relative relation graph is introduced to guide attention weight allocation, forming a Relation-Guided Transformer (RGTs) that dynamically modulate attention weights based on relative attribute relations to capture the interplay between semantics and various social attributes. The refined features are fused with the target instance for popularity prediction. Experiments on three public benchmarks show that RE-Rag consistently outperforms state-of-the-art methods in both prediction accuracy and retrieval efficiency.

Figures

Figures reproduced from arXiv: 2607.19200 by An-An Liu, Bolun Zheng, Canjin Wang, Lingyu Zhu, Ning Xu, Wei Zhu, Yuhao Luo.

Figure 1
Figure 1. Figure 1: Comparison among (a) non-retrieval models, (b) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed RE-Rag framework: (a) Semantic-Attribute Retriever combines multimodal semantic [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Average popularity gap between query samples and [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Parameter sensitivity analysis of RE-Rag on three [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Impact of training set proportion on SMPD: bars [PITH_FULL_IMAGE:figures/full_fig_p008_5.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

52 extracted references · 5 linked inside Pith

  1. [1]

    Fatma S Abousaleh, Wen-Huang Cheng, Neng-Hao Yu, and Yu Tsao. 2020. Multi- modal deep learning framework for image popularity prediction on social media. IEEE Transactions on Cognitive and Developmental Systems13, 3 (2020), 679–692

    2020

  2. [2]

    Nesrine Ben Hassine, Pascale Minet, Dana Marinca, and Dominique Barth. 2019. Popularity prediction–based caching in content delivery networks.Annals of Telecommunications74, 5 (2019), 351–364

    2019

  3. [3]

    Sebastian Borgeaud and et al. 2022. Improving language models by retrieving from trillions of tokens. InICML

    2022

  4. [4]

    Jingyuan Chen, Xuemeng Song, Liqiang Nie, Xiang Wang, Hanwang Zhang, and Tat-Seng Chua. 2016. Micro tells macro: Predicting the popularity of micro-videos via a transductive model. InProceedings of the 24th ACM international conference on Multimedia. 898–907

    2016

  5. [5]

    Justin Cheng, Lada Adamic, P Alex Dow, Jon Michael Kleinberg, and Jure Leskovec. 2014. Can cascades be predicted?. InProceedings of the 23rd inter- national conference on World wide web. 925–936

    2014

  6. [6]

    Zhangtao Cheng, Jiao Li, Jian Lang, Ting Zhong, and Fan Zhou. 2025. In-context Prompt-augmented Micro-video Popularity Prediction. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 11527–11535

    2025

  7. [7]

    Zhangtao Cheng, Jienan Zhang, Xovee Xu, Goce Trajcevski, Ting Zhong, and Fan Zhou. 2024. Retrieval-augmented hypergraph for multimodal social media popularity prediction. InProceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining. 445–455

    2024

  8. [8]

    Tsun-hin Cheung and Kin-man Lam. 2022. Crossmodal bipolar attention for multimodal classification on social media.Neurocomputing514 (2022), 1–12

    2022

  9. [9]

    Minhwa Cho, Dahye Jeong, and Eunil Park. 2024. AMPS: Predicting popularity of short-form videos using multi-modal attention mechanisms in social media marketing environments.Journal of Retailing and Consumer Services78 (2024), 103778

    2024

  10. [10]

    Birds of a Feather

    Munmun De Choudhury, Hari Sundaram, Ajita John, Doree Duncan Seligmann, and Aisling Kelliher. 2010. " Birds of a Feather": Does User Homophily Impact Information Diffusion in Social Media?arXiv preprint arXiv:1006.1702(2010)

    Pith/arXiv arXiv 2010

  11. [11]

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yixin Dai, Jiawei Sun, Haofen Wang, Haofen Wang, et al. 2023. Retrieval-augmented generation for large language models: A survey.arXiv preprint arXiv:2312.10997 2, 1 (2023), 32

    Pith/arXiv arXiv 2023

  12. [12]

    Chih-Chung Hsu, Chia-Ming Lee, Xiu-Yu Hou, and Chi-Han Tsai. 2023. Gradient boost tree network based on extensive feature analysis for popularity prediction of social posts. InProceedings of the 31st ACM International Conference on Multimedia. 9451–9455

    2023

  13. [13]

    Liya Ji, Chan Ho Park, Zhefan Rao, and Qifeng Chen. 2023. Neural image pop- ularity assessment with retrieval-augmented transformer. InProceedings of the 31st ACM International Conference on Multimedia. 2427–2436

    2023

  14. [14]

    Xinke Jiang, Rihong Qiu, Yongxin Xu, Yichen Zhu, Ruizhe Zhang, Yuchen Fang, Chu Xu, Junfeng Zhao, and Yasha Wang. 2024. Ragraph: A general retrieval- augmented graph learning framework.Advances in Neural Information Processing Systems37 (2024), 29948–29985

    2024

  15. [16]

    Aditya Khosla, Atish Das Sarma, and Raffay Hamid. 2014. What makes an image popular?. InProceedings of the 23rd international conference on World wide web. 867–876

    2014

  16. [17]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2017. Adam: A Method for Stochastic Opti- mization. arXiv:1412.6980 [cs.LG] https://arxiv.org/abs/1412.6980

    Pith/arXiv arXiv 2017

  17. [18]

    Xin Lai, Yihong Zhang, and Wei Zhang. 2020. HyFea: Winning solution to social media popularity prediction for multimedia grand challenge 2020. InProceedings of the 28th ACM International Conference on Multimedia. 4565–4569

    2020

  18. [19]

    Patrick Lewis and et al. 2020. Retrieval-Augmented Generation for Knowledge- Intensive NLP Tasks. InNeurIPS

    2020

  19. [20]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems33 (2020), 9459–9474

    2020

  20. [21]

    Chen Li, Xiaoyu Wang, Tongyu Zong, Houwei Cao, and Yong Liu. 2023. Predic- tive edge caching through deep mining of sequential patterns in user content retrievals.Computer Networks233 (2023), 109866

    2023

  21. [22]

    Dongliang Liao, Jin Xu, Gongfu Li, Weijie Huang, Weiqing Liu, and Jing Li. 2019. Popularity prediction on online articles with deep fusion of temporal process and content features. InProceedings of the AAAI conference on artificial intelligence, Vol. 33. 200–207

    2019

  22. [23]

    Yin Luo, Fangfang Wang, Feifei Zhao, Jianbin Guo, Lei Wang, Yanni Hao, and Daniel Dajun Zeng. 2019. A framework for policy information popularity pre- diction in new media. In2019 IEEE International Conference on Intelligence and Security Informatics (ISI). IEEE, 209–211

    2019

  23. [24]

    Jinna Lv, Wu Liu, Meng Zhang, He Gong, Bin Wu, and Huadong Ma. 2017. Multi- feature fusion for predicting social media popularity. InProceedings of the 25th ACM international conference on Multimedia. 1883–1888

    2017

  24. [25]

    Mira Mayrhofer, Jörg Matthes, Sabine Einwiller, and Brigitte Naderer. 2020. User generated content presenting brands on social media increases young adults’ purchase intention.International Journal of Advertising39, 1 (2020), 166–186

    2020

  25. [26]

    2025.Digital news report 2025

    Nic Newman, Arguedas Ross Arguedas, Craig T Robertson, Rasmus Kleis Nielsen, and Richard Fletcher. 2025.Digital news report 2025. Reuters Institute for the study of Journalism

    2025

  26. [27]

    Alessandro Ortis, Giovanni Maria Farinella, and Sebastiano Battiato. 2019. Pre- diction of social image popularity dynamics. InInternational Conference on Image Analysis and Processing. Springer, 572–582

    2019

  27. [28]

    Boci Peng, Yun Zhu, Yongchao Liu, Xiaohe Bo, Haizhou Shi, Chuntao Hong, Yan Zhang, and Siliang Tang. 2024. Graph retrieval-augmented generation: A survey. arXiv preprint arXiv:2408.08921(2024)

    Pith/arXiv arXiv 2024

  28. [29]

    Henrique Pinto, Jussara M Almeida, and Marcos A Gonçalves. 2013. Using early view patterns to predict the popularity of youtube videos. InProceedings of the sixth ACM international conference on Web search and data mining. 365–374

    2013

  29. [31]

    Yang Qian, Wang Xu, Xiao Liu, Haifeng Ling, Yuanchun Jiang, Yidong Chai, and Yezheng Liu. 2022. Popularity prediction for marketer-generated content: A text-guided attention neural network for multi-modal feature fusion.Information Processing & Management59, 4 (2022), 102984

    2022

  30. [32]

    Yingdan Shang, Bin Zhou, Xiang Zeng, Ye Wang, Han Yu, and Zhong Zhang

  31. [33]

    Fabian Spaeh and Alina Ene. 2023. Online ad allocation with predictions.Advances in Neural Information Processing Systems36 (2023), 17265–17295

    2023

  32. [34]

    Karen Sparck Jones. 1972. A statistical interpretation of term specificity and its application in retrieval.Journal of documentation28, 1 (1972), 11–21

    1972

  33. [35]

    Mohd Suhairi Md Suhaimin, Mohd Hanafi Ahmad Hijazi, Ervin Gubin Moung, Puteri Nor Ellyza Nohuddin, Stephanie Chua, and Frans Coenen. 2023. Social media sentiment analysis and opinion mining in public security: Taxonomy, trend analysis, issues and future directions.Journal of King Saud University-Computer and Information Sciences35, 9 (2023), 101776

    2023

  34. [36]

    Dachun Sun, You Lyu, Jinning Li, Yizhuo Chen, Tianshi Wang, Tomoyoshi Kimura, and Tarek Abdelzaher. 2025. SCRAG: Social Computing-Based Retrieval Aug- mented Generation for Community Response Forecasting in Social Media Envi- ronments. In2025 IEEE International Conference on Smart Computing (SMART- COMP). IEEE, 170–177

    2025

  35. [37]

    Gabor Szabo and Bernardo A Huberman. 2010. Predicting the popularity of online content.Commun. ACM53, 8 (2010), 80–88

    2010

  36. [38]

    Alexandru Tatar, Marcelo Dias De Amorim, Serge Fdida, and Panayotis Anto- niadis. 2014. A survey on predicting the popularity of web content.Journal of Internet Services and Applications5, 1 (2014), 8

    2014

  37. [39]

    Jing Tian, Huayin Fan, and Zengwen Hou. 2022. Research on the prediction of popularity of news dissemination public opinion based on data mining.Compu- tational Intelligence and Neuroscience2022, 1 (2022), 6512602

    2022

  38. [40]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in Neural Information Processing Systems30 (2017)

    2017

  39. [41]

    Benyou Wang, Lifeng Shang, Christina Lioma, Xin Jiang, Hao Yang, Qun Liu, and Jakob Grue Simonsen. 2020. On position embeddings in bert. InInternational conference on learning representations

    2020

  40. [42]

    Haoyu Wang, Zongxia Xie, Meiyao Liu, and Canhua Guan. 2023. Afrf: angle feature retrieval based popularity forecasting. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management. 2606–2615

    2023

  41. [43]

    Evan Weissburg, Arya Kumar, and Paramveer S Dhillon. 2022. Judging a book by its cover: Predicting the marginal impact of title on Reddit post popularity. In Proceedings of the International AAAI Conference on Web and Social Media, Vol. 16. 1098–1108

    2022

  42. [44]

    Bo Wu, Wen-Huang Cheng, Yongdong Zhang, Qiushi Huang, Jintao Li, and Tao Mei. 2017. Sequential prediction of social media popularity with deep temporal context networks.arXiv preprint arXiv:1712.04443(2017)

    Pith/arXiv arXiv 2017

  43. [45]

    Bo Wu, Peiye Liu, Qiushi Huang, Zhaoyang Zeng, Jia Wang, Bei Liu, Jiebo Luo, and Wen-Huang Cheng. 2024. SMP Challenge Summary: Social Media Prediction Challenge. InProceedings of the 32nd ACM International Conference on Multimedia. 11442–11444

    2024

  44. [46]

    Max Würfel, Qiwei Han, and Maximilian Kaiser. 2021. Online advertising revenue forecasting: An interpretable deep learning approach. In2021 IEEE International Conference on Big Data (Big Data). IEEE, 1980–1989

    2021

  45. [47]

    Jiayi Xie, Yaochen Zhu, and Zhenzhong Chen. 2021. Micro-video popularity prediction via multimodal variational information bottleneck.IEEE Transactions on Multimedia25 (2021), 24–37

    2021

  46. [48]

    Xovee Xu, Yifan Zhang, Fan Zhou, and Jingkuan Song. 2025. Improving Mul- timodal Social Media Popularity Prediction via Selective Retrieval Knowledge Augmentation. InProceedings of the AAAI Conference on Artificial Intelligence, Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Trovato et al. Vol. 39. 932–940

    2025

  47. [49]

    Yijie Xu, Bolun Zheng, Wei Zhu, Hangjia Pan, Yuchen Yao, Ning Xu, Anan Liu, Quan Zhang, and Chenggang Yan. 2025. SMTPD: A New Benchmark for Temporal Prediction of Social Media Popularity. InProceedings of the Computer Vision and Pattern Recognition Conference. 18847–18857

    2025

  48. [50]

    Wei Zhang, Wen Wang, Jun Wang, and Hongyuan Zha. 2018. User-guided hierarchical attention network for multi-modal social image popularity prediction. InProceedings of the 2018 world wide web conference. 1277–1286

    2018

  49. [51]

    Zhang and et al

    Y. Zhang and et al. 2021. RA-Rec: A Retrieval-Augmented Recommendation Framework. InSIGIR

    2021

  50. [52]

    Zhuoran Zhang, Shibiao Xu, Li Guo, and Wenke Lian. 2022. Multi-modal Varia- tional Auto-Encoder Model for Micro-video Popularity Prediction. InProceedings of the 8th International Conference on Communication and Information Processing. 9–16

    2022

  51. [53]

    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

  52. [2022]

    Predicting the popularity of online content by modeling the social influence and homophily features.Frontiers in Physics10 (2022), 915756

    2022