Pith. sign in

REVIEW 4 major objections 7 minor 39 references

HyperFusion: Hierarchical Multimodal Ensemble Learning for Social Media Popularity Prediction

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A three-tier fusion of image, text, and user data predicts social media popularity with rank correlation of 0.7324.

desk verdict A competent challenge entry whose headline 'hierarchical fusion' never appears in the method; the paper needs major revision or an honest reframing before publication. read the letter →

arxiv 2507.00926 v1 pith:GKS4AXID submitted 2025-07-01 cs.MM cs.LG

Liliang Ye (1) , Yunyao Zhang (1) , Yafeng Wu (1) , Yi-Ping Phoebe Chen (2) , Junqing Yu (1) , Wei Yang (1) , Zikai Song (1) ((1) Huazhong University of Science , Technology
show 5 more authors
Wuhan China (2) La Trobe University Melbourne Australia)
This is my paper
classification cs.MMcs.LG
keywords SocialmediapopularitypredictionMultimodalfusionHierarchicalarchitectureEnsemblelearningPseudo-labelingCLIPembeddingsCatBoostSMPChallenge
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 argues that predicting how many people will engage with a social media post is best done by fusing information in stages, not by throwing all features into one vector. It introduces HyperFusion, a three-tier architecture that combines CLIP-based visual and textual embeddings, spatiotemporal and user metadata, SVD-derived latent embeddings, and a cross-modal similarity score, and then feeds the assembled features to an ensemble of CatBoost, TabNet, and a custom MLP. On the SMP challenge validation set the system reports a Spearman rank correlation (SRC) of 0.7324 and a mean absolute error (MAE) of 1.2402, which the authors state placed their entry third in the SMP 2025 Image Track. The paper's central claim is that the hierarchical fusion design and the iterative pseudo-labeling training stage are what deliver that accuracy.

What carries the argument

The load-bearing mechanism is the three-tier hierarchical fusion architecture, a pipeline that integrates features at increasing abstraction levels instead of a single flat concatenation. In the authors' design, tier one extracts per-modality representations (CLIP visual, CLIP text plus GloVe, statistical text features, spatiotemporal and user metadata); tier two produces latent user and location embeddings via SVD and a cross-modal similarity score from CLIP image-text cosine similarity; tier three sends the unified feature vector through an ensemble of CatBoost, TabNet, an MLP with co-attention, and a CLIP-based hierarchical predictor. The training machinery is a two-stage semi-supervised procedure that trains base models on labeled data, then thresholds pseudo-labels by $\tau = \mu_{\mathrm{confidence}} + \alpha \sigma_{\mathrm{confidence}}$ and retrains iteratively. Five-fold cross-validation with optimized weights $w_i^{(k)}$ combines the models into the final prediction.

What would settle it

Run the same feature set through a single flat concatenation (or a standard non-hierarchical fusion) with the same models, folds, and pseudo-labeling on the SMP validation set; if that baseline reaches or exceeds SRC of about 0.7324, the paper's central architectural claim is not supported. A second check is whether user ID remains the top feature when predictions are made for users never seen in training.

Watch

Extended reading notes

Core claim

The central claim is that hierarchical multimodal fusion is the right structure for social media popularity prediction. HyperFusion first extracts visual vectors from a CLIP image encoder, textual vectors from CLIP text and GloVe embeddings, and statistical text features; it then derives spatiotemporal and user features, applies SVD to interaction matrices to get user and location embeddings, and computes a CLIP cosine similarity between image and text as a coherence measure. All features are concatenated and passed through an ensemble of CatBoost, TabNet, and a multi-layer perceptron, with a five-fold cross-validated weighted average as the final predictor. A two-stage training loop adds high-confidence pseudo-labels from the test set and retrains. The authors report that the full system reaches SRC 0.7324 and MAE 1.2402, and that removing visual features, metadata, textual features, the ensemble, pseudo-labeling, filtering, or five-fold validation each degrades performance, with visual features causing the largest drop.

Load-bearing premise

The claim that the hierarchical fusion architecture is responsible for the reported accuracy assumes that the ablations isolate the fusion design; they do not, because no flat-concatenation or prior-method baseline is reported.

Editorial extensions

If this is right

  • If visual features are ablated, SRC drops from 0.7324 to 0.6916, so image content is the single largest modality contribution in this system.
  • The pseudo-labeling stage raises SRC by roughly 0.004 over the same pipeline without it, indicating unlabeled test posts are worth exploiting when labels are scarce.
  • User-centric features such as user ID, post count, and follower count occupy the top ranks of feature importance, so behavioral metadata is carrying much of the predictive signal.
  • The five-fold ensemble and outlier filtering each add measurable SRC, suggesting the final score depends on training-detail choices as much as on fusion structure.

Reading between the lines

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

  • The paper's ablations remove whole modality groups or the ensemble, but never compare flat concatenation of the same features to the three-tier pipeline; that comparison is needed to attribute the gain to hierarchy rather than to feature engineering alone.
  • Because user ID is the top-ranked feature, much of the signal may come from memorizing user-specific posting habits; splitting the evaluation by held-out users would test whether the method generalizes to unseen accounts.
  • The cross-modal coherence measure is a single scalar cosine score; a natural extension is a region-aware or attention-weighted alignment (for example, CLIP patch-text similarity), which the paper does not explore.
  • The pseudo-label threshold depends on the mean and spread of ensemble confidence; an alternative threshold based on ensemble agreement (variance across the component models) could be tested against the paper's choice.
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 / 7 minor

Summary. The manuscript describes HyperFusion, a system for social media popularity prediction on the SMP 2025 Image Track, and reports a Spearman rank correlation of 0.7324 and MAE of 1.2402, with the team placed third. The proposed pipeline extracts CLIP visual features, CLIP/GloVe/text-statistics textual features, spatiotemporal and user features, and a CLIP cosine cross-modal similarity, concatenates them (Eq. 6), and feeds the vector to CatBoost, TabNet, MLP, and a CLIP-based predictor whose outputs are weighted-averaged (Eqs. 8-9); a two-stage pseudo-labeling procedure uses high-confidence ensemble predictions (Eq. 10). The paper claims a three-tier hierarchical fusion architecture and hierarchical clustering features, and presents ablations (Table 1) removing modalities, the ensemble, pseudo-labeling, filtering, and 5-fold.

Significance. If the hierarchical-fusion claim were substantiated, the paper would provide a competitive challenge solution and a useful study of modality contributions on the SMP dataset. The reported numbers are plausible and the ablation table covers several pipeline components; the paper's availability of code is a plus. However, the central novelty is not present in the formal method: the fusion described by Eq. 6 is flat concatenation and Eq. 8 is a weighted average, with no tiered intermediate fusion, and hierarchical clustering features are never defined. Because the key method is missing, the manuscript's main architectural claim is currently not testable, and the experimental section does not compare against the flat-concatenation baseline it claims to beat in Section 1.

major comments (4)
  1. [§3.3–3.4, Eqs. (6) and (8)] The claimed three-tier hierarchical fusion architecture is not described. Eq. (6) concatenates all modality vectors into a single vector, and Eq. (8) is a weighted average of model outputs; no equation or pseudocode defines progressive, tiered integration, and no model input depends on another model's output. As written, the method is a flat-concatenation plus weighted-average ensemble, so the central claim in the abstract and Section 1 is unsupported. Please either provide the full hierarchy (stage-by-stage equations, intermediate outputs, tier structure) or revise the claims to match the implemented method.
  2. [Abstract and §3.3] The paper promises 'hierarchical clustering features' that capture inter-modal dependencies, but no such features are defined in the methodology, appear in Eq. (6), or show up in the feature-importance analysis of Table 2. Please specify how the clustering is performed, at which hierarchy level features are produced, and include them in the ablations; otherwise remove the claim.
  3. [§4.3.2, Table 1] The ablation study removes whole modality groups (visual, textual, metadata, geolocation) or entire components (ensemble, pseudo-label, filter, 5-fold), but it never removes or varies the hierarchical structure itself, and it never compares against a flat-concatenation baseline or prior published methods. Consequently, the sentence in Section 1 claiming that 'the hierarchical fusion approach consistently outperforms traditional concatenation methods' has no experimental support in this manuscript. Please add such baselines (e.g., flat concatenation with a single model and with the same ensemble, and at least one published prior method) and report the comparison.
  4. [§4.3.1 and §4.3.2] All reported SRC/MAE values appear to be single-run results without error bars or significance tests. Several differences are tiny (e.g., w/o Filter SRC 0.7249 vs. full 0.7324, w/o Geo 0.7294 vs. 0.7324), so it is not possible to tell whether the differences are meaningful. Please report mean and standard deviation over multiple runs or over the five folds, and state whether the validation-set numbers are from the same split used for model selection.
minor comments (7)
  1. [§4.2 and Table 1] The text refers to the 'official evaluation dataset' while Table 1's caption says 'validation set'; please clarify which set each number comes from.
  2. [Eq. (10)] Equation (10) defines a threshold using 'prediction confidence,' but the manuscript does not define confidence for a regression ensemble; please define it explicitly.
  3. [Figure 1] Figure 1 labels 'Feature transformer Step 1/Step 2 FC' but the text never describes these components; either add the description or remove the labels.
  4. [References] The reference list is duplicated verbatim at the end of the manuscript; please keep a single bibliography.
  5. [§3.4–3.5] Hyperparameters such as ensemble weights w_i^(k), threshold alpha, and Huber delta are not reported; please provide values or a link to a configuration file for reproducibility.
  6. [§4.3.3 and Table 2] Table 2 does not list any cross-modal similarity or hierarchical clustering feature, despite the text in Section 4.3.3 saying such measures contribute; please reconcile.
  7. [Abstract and §1] The claim that 'our team achieved third place' should cite the official leaderboard or challenge report, and the anonymous code link should be replaced with a permanent repository.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; holdout evaluation and transductive pseudo-labeling are self-contained, while the under-specified 'hierarchical' architecture is a correctness gap, not circularity.

full rationale

The reported SRC=0.7324 and MAE=1.2402 come from evaluating the trained ensemble on the official evaluation/validation set; no parameter is fit to those labels and then renamed as a prediction. The pseudo-labeling stage (Section 3.5) uses the model's own high-confidence test predictions as extra training targets, a transductive semi-supervised technique that can help or hurt but does not by construction force the reported numbers. The ensemble weights in Eq. 8 are tuned on validation data, not on the evaluation labels, and the five-fold averaging in Eq. 9 is a standard aggregation. No load-bearing claim rests on a self-citation: the overlapping-author references ([10,15,19,25,26,27,28,39]) are background tracking/video papers and are not invoked as a uniqueness theorem or as justification for the main result. The genuine weakness is that the abstract's 'three-tier fusion' and 'hierarchical clustering features' are not realized in Eq. 6 (flat concatenation) or Eq. 8 (weighted average), and the ablations never compare flat concatenation or prior methods; that is an internal-consistency/evidence problem for the architectural claim, not a circular derivation. I therefore find no circularity and score 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The central result depends on many hand-set choices and unspecified constructs. No parameter values are reported for the pseudo-label threshold, Huber loss, ensemble weights, or base-model hyperparameters, and the hierarchical clustering features are undefined. This makes the contribution more of an engineering configuration than a derived method.

free parameters (5)
  • Ensemble weights w_i^(k) = Optimized per fold via cross-validation
    Eq. 8 introduces per-fold weighted averaging; the optimal weights are fitted to validation predictions, not derived from first principles.
  • Pseudo-label threshold alpha = Not reported
    Eq. 10 controls how many pseudo-labels are kept; the value is empirically determined, but no value or search range is reported.
  • Huber loss delta = Not reported
    Eq. 7 defines the transition point between quadratic and linear loss; the value is not reported.
  • SVD embedding dimensions = 399 for users, 400 for locations
    Section 3.3 chooses these ranks for user and location embeddings with no stated criterion.
  • Unreported base-model hyperparameters = Not reported
    CatBoost, TabNet, and MLP settings such as depth, learning rate, and layer sizes are not listed, so replication requires guessing them.
assumptions (4)
  • domain assumption Pretrained CLIP and GloVe embeddings transfer semantic meaning to popularity prediction.
    Section 3.3 uses these features as inputs; their validity for this task is assumed, not demonstrated.
  • domain assumption SVD of user-item interactions captures latent user and location behavior relevant to popularity.
    Section 3.3, Eq. 4 assumes that low-rank user and location embeddings encode predictive patterns; no evaluation of embedding quality is given.
  • domain assumption High-confidence pseudo-labels from the model's own predictions improve generalization.
    Section 3.5, Eq. 10 assumes that the threshold keeps only reliable labels; no validation of pseudo-label accuracy is provided.
  • standard math Standard math tools (PCA, SVD, Huber loss, cosine similarity) are valid for the stated preprocessing.
    These tools are invoked in Eqs. 2 through 7 as background mathematics without proof.
invented entities (1)
  • Hierarchical clustering features
    purpose: Claimed in the abstract and introduction to capture inter-modal dependencies.
    The term appears in Section 1 and is mentioned as part of the feature set, but no clustering algorithm, feature formulation, or ablation entry is provided anywhere in the paper. There is no way to measure its contribution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HyperFusion: Hierarchical Multimodal Ensemble Learning for Social Media Popularity Prediction." pith.science (2026). https://pith.science/paper/GKS4AXID

@misc{pith2026250700926,
  author       = {Pith},
  title        = {Pith review of: HyperFusion: Hierarchical Multimodal Ensemble Learning for Social Media Popularity Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GKS4AXID}},
  note         = {Machine review of arXiv:2507.00926}
}
read the original abstract

Social media popularity prediction plays a crucial role in content optimization, marketing strategies, and user engagement enhancement across digital platforms. However, predicting post popularity remains challenging due to the complex interplay between visual, textual, temporal, and user behavioral factors. This paper presents HyperFusion, a hierarchical multimodal ensemble learning framework for social media popularity prediction. Our approach employs a three-tier fusion architecture that progressively integrates features across abstraction levels: visual representations from CLIP encoders, textual embeddings from transformer models, and temporal-spatial metadata with user characteristics. The framework implements a hierarchical ensemble strategy combining CatBoost, TabNet, and custom multi-layer perceptrons. To address limited labeled data, we propose a two-stage training methodology with pseudo-labeling and iterative refinement. We introduce novel cross-modal similarity measures and hierarchical clustering features that capture inter-modal dependencies. Experimental results demonstrate that HyperFusion achieves competitive performance on the SMP challenge dataset. Our team achieved third place in the SMP Challenge 2025 (Image Track). The source code is available at https://anonymous.4open.science/r/SMPDImage.

Figures

Figures reproduced from arXiv: 2507.00926 by the authors.

Figure 1
Figure 1. The HyperFusion framework pipeline integrates visual, textual, spatiotemporal, and user features through hierarchical [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Histogram and kernel density estimation of pre [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 15 canonical work pages

  1. [1]

    Sercan Ö Arik and Tomas Pfister. 2021. Tabnet: Attentive interpretable tabular learning. In Proceedings of the AAAI conference on artificial intelligence , Vol. 35. 6679–6687

    work page 2021

  2. [2]

    Tianqi Chen and Carlos Guestrin. 2016. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining . 785–794

    2016

  3. [3]

    Xiaolu Chen, https://orcid.org/0000-0001-5819-5371, View Profile, Wei- long Chen, https://orcid.org/0000-0003-2202-601X, View Profile, Chenghao Huang, https://orcid.org/0000-0003-2961-6282, View Profile, Zhongjian Zhang, https://orcid.org/0009-0000-6180-4342, View Profile, Lixin Duan, https://orcid.org/0000-0002-0723-4016, View Profile, Yanru Zhang, https...

    arXiv 2023

  4. [4]

    Zhangtao Cheng, Jienan Zhang, Xovee Xu, Goce Trajcevski, Ting Zhong, and Fan Zhou. 2024. Retrieval-Augmented Hypergraph for Multimodal Social Media Popularity Prediction. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’24) . Association for Computing Machinery, New York, NY, USA, 445–455. doi:10.1145/3637528.3672041

    arXiv 2024

  5. [5]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers). 4171–4186

    2019

  6. [6]

    Anna Veronika Dorogush, Vasily Ershov, and Andrey Gulin. 2018. CatBoost: gra- dient boosting with categorical features support. arXiv preprint arXiv:1810.11363 (2018)

    arXiv 2018

  7. [7]

    Harold Hotelling. 1933. Analysis of a complex of statistical variables into principal components. Journal of educational psychology 24, 6 (1933), 417

    work page 1933

  8. [8]

    Chih-Chung Hsu, Chia-Ming Lee, Yu-Fan Lin, Yi-Shiuan Chou, Chih-Yu Jian, and Chi-Han Tsai. 2024. Revisiting Vision-Language Features Adaptation and Inconsistency for Social Media Popularity Prediction. arXiv:2407.00556 doi:10. 48550/arXiv.2407.00556

    work page Pith review arXiv doi:10.48550/arxiv.2407.00556 2024

Show all 39 references
  1. [10]

    Yangliu Hu, Zikai Song, Na Feng, Yawei Luo, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. 2025. Sf2t: Self-supervised fragment finetuning of video-llms for fine-grained understanding. In Proceedings of the Computer Vision and Pattern Recognition Conference. 29108–29117

    2025

  2. [11]

    Dahyun Jeong, Hyelim Son, Yunjin Choi, and Keunwoo Kim. 2024. Enhancing social media post popularity prediction with visual content. Journal of the Korean Statistical Society 53, 3 (2024), 844–882

    2024

  3. [12]

    Peipei Kang, Zehang Lin, Shaohua Teng, Guipeng Zhang, Lingni Guo, and Wei Zhang. 2019. Catboost-based framework with additional user information for social media popularity prediction. In Proceedings of the 27th ACM international conference on multimedia. 2677–2681

    2019

  4. [13]

    Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. 2017. Lightgbm: A highly efficient gradient boosting MM ’25, October 27–31, 2025, Dublin, Ireland Liliang Ye et al. decision tree. Advances in neural information processing system...

    2017

  5. [14]

    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 (MM ’20) . Association for Computing Machinery, New York, NY, USA...

    2020 doi

  6. [15]

    Wenbing Li, Hang Zhou, Junqing Yu, Zikai Song, and Wei Yang. 2024. Coupled mamba: Enhanced multi-modal fusion with coupled state space model. arXiv preprint arXiv:2405.18014 (2024)

    2024 arXiv

  7. [16]

    Hung-Hsiang Lin, Jiun-Da Lin, Jose Jaena Mari Ople, Jun-Cheng Chen, and Kai-Lung Hua. 2021. Social media popularity prediction based on multi-modal self-attention mechanisms. IEEE Access 10 (2021), 4448–4455

    2021

  8. [17]

    Yu-Shi Lin and Anthony J.T. Lee. 2024. MMF: Winning Solution to Social Media Popularity Prediction Challenge 2024. In Proceedings of the 32nd ACM Interna- tional Conference on Multimedia (MM ’24). Association for Computing Machinery, New York, NY, USA, 11445–11449. doi:10.1145...

    2024

  9. [18]

    Yu-Shi Lin and Anthony JT Lee. 2024. MMF: Winning Solution to Social Media Popularity Prediction Challenge 2024. In Proceedings of the 32nd ACM Interna- tional Conference on Multimedia . 11445–11449

    2024

  10. [19]

    Run Luo, Zikai Song, Lintao Ma, Jinlin Wei, Wei Yang, and Min Yang. 2024. Diffusiontrack: Diffusion model for multi-object tracking. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 38. 3991–3999

    2024

  11. [20]

    Shijian Mao, Wudong Xi, Lei Yu, Gaotian Lü, Xingxing Xing, Xingchen Zhou, and Wei Wan. 2023. Enhanced catboost with stacking features for social media prediction. In Proceedings of the 31st ACM International Conference on Multimedia . 9430–9435

    2023

  12. [21]

    Mayank Meghawat, Satyendra Yadav, Debanjan Mahata, Yifang Yin, Rajiv Ratn Shah, and Roger Zimmermann. 2018. A multimodal approach to predict social media popularity. In 2018 IEEE conference on multimedia information processing and retrieval (MIPR). IEEE, 190–195

    2018

  13. [22]

    Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP) . 1532–1543

    2014

  14. [23]

    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

  15. [24]

    Social Media Statistic [n. d.]. Global Social Media Statistics. https://datareportal.com/social-media-users

  16. [25]

    Zikai Song, Run Luo, Lintao Ma, Ying Tang, Yi-Ping Phoebe Chen, Junqing Yu, and Wei Yang. 2025. Temporal Coherent Object Flow for Multi-Object Tracking. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 6978–6986

    2025

  17. [26]

    Zikai Song, Run Luo, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. 2023. Compact transformer tracker with correlative masked modeling. In Proceedings of the AAAI conference on artificial intelligence , Vol. 37. 2321–2329

    2023

  18. [27]

    Zikai Song, Ying Tang, Run Luo, Lintao Ma, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. 2024. Autogenic language embedding for coherent point tracking. In Proceedings of the 32nd ACM International Conference on Multimedia . 2021–2030

    2024

  19. [28]

    Zikai Song, Junqing Yu, Yi-Ping Phoebe Chen, and Wei Yang. 2022. Transformer tracking with cyclic shifting window attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 8791–8800

    2022

  20. [29]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems 30 (2017)

    2017

  21. [30]

    Jing Wang, Shuo Yang, Hui Zhao, and Yue Yang. 2023. Social media popular- ity prediction with multimodal hierarchical fusion model. Computer Speech & Language 80 (2023), 101490

    2023

  22. [31]

    Kai Wang, Penghui Wang, Xin Chen, Qiushi Huang, Zhendong Mao, and Yong- dong Zhang. 2020. A Feature Generalization Framework for Social Media Pop- ularity Prediction. In Proceedings of the 28th ACM International Conference on Multimedia (MM ’20). Association for Computing Mach...

    2020

  23. [32]

    Bo Wu, Wen-Huang Cheng, Peiye Liu, Bei Liu, Zhaoyang Zeng, and Jiebo Luo

  24. [33]

    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:1712.04443 [cs] doi:10.48550/arXiv.1712.04443

    2017 doi arXiv

  25. [34]

    Bo Wu, Peiye Liu, Wen-Huang Cheng, Bei Liu, Zhaoyang Zeng, Jia Wang, Qiushi Huang, and Jiebo Luo. 2023. SMP Challenge: An overview and analysis of social media prediction challenge. In Proceedings of the 31st ACM International Conference on Multimedia. 9651–9655

    2023

  26. [35]

    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. In Proceedings of the 32nd ACM International Conference on Multimedia. 11442–11444

    2024

  27. [36]

    Bo Wu, Tao Mei, Wen-Huang Cheng, and Yongdong Zhang. 2016. Unfolding temporal dynamics: Predicting social media popularity using multi-scale temporal decomposition. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 30

    2016

  28. [37]

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

    2018

  29. [38]

    Zhizhen Zhang, Ruihong Qiu, and Xiaohui Xie. 2024. Contrastive Learn- ing for Implicit Social Factors in Social Media Popularity Prediction. https://arxiv.org/abs/2410.09345v1

    2024 arXiv

  30. [39]

    Hang Zhou, Jiale Cai, Yuteng Ye, Yonghui Feng, Chenxing Gao, Junqing Yu, Zikai Song, and Wei Yang. 2025. Video anomaly detection with motion and appearance guided patch diffusion model. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 10761–10769

    2025

  31. [2019]

    In Proceedings of the 27th ACM International Conference on Multimedia

    Smp challenge: An overview of social media prediction challenge 2019. In Proceedings of the 27th ACM International Conference on Multimedia . 2667–2671

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.