Pith. sign in

REVIEW 2 major objections 6 minor 1 cited by

Synergizing LLMs with Global Label Propagation for Multimodal Fake News Detection

T0 review · 2 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that spreading LLM-generated pseudo labels over a cross-modal tweet graph, trained with a random-mask rule, beats state-of-the-art multimodal fake news detectors on Twitter, PHEME, and Weibo.

desk verdict Clear combination of known pieces with a new training mask, but inference-time self-pseudo-label concatenation and test-set hyperparameter tuning undermine the claim that label propagation drives the gains. read the letter →

arxiv 2506.00488 v1 pith:L3JIPGW4 submitted 2025-05-31 cs.CL

Shuguo Hu , Jun Hu , Huaiwen Zhang This is my paper · ORCID
classification cs.CL
keywords fakenewsdetectionmultimodallabelpropagationlargelanguagemodelspseudolabelinggraphneuralnetworksglobalrandommaskcross-modal
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

The paper argues that large language model predictions, which are too weak to use on their own, become a decisive advantage when fed into a global label propagation network as pseudo labels. It proposes GLPN-LLM, which builds a graph over news items from text and image similarity, attaches one-hot label embeddings (true labels where available, LLM pseudo labels where confident) to each node, and runs a GCN over the graph. A Global Random Mask zeros the label embedding of a random subset of nodes during training and computes the loss only on those nodes, so the model learns to propagate neighbors' label information instead of copying a node's own label. At inference the test node's own pseudo label is included in its features, and the paper reports consistent gains over all baselines, including a larger F1 increase than simply adding LLM outputs to the prior label-propagation baseline.

What carries the argument

The load-bearing mechanism is the Global Random Mask (GRM) coupled with label-integrated node features. During training, a fraction $\rho$ (best at 0.5) of nodes have their label embedding $\tilde{y}_i$ zeroed via a binary mask and serve as the only loss nodes, so the GCN must infer their labels from neighboring nodes' content and label features; at inference every test node's own LLM pseudo label is restored into its input via $x'_i = x_i \oplus y'_i$. This trains the propagation to use label information flowing across the graph while preventing a node from trivially copying its own label, and it is what lets weak LLM guesses be collectively corrected.

What would settle it

Run GLPN-LLM at inference with all test pseudo labels zeroed so that no node sees its own label feature; if F1 drops to near the FCN-LP baseline, the gains come from the test node's own label rather than global propagation. Alternatively, shuffle the LLM pseudo labels among test nodes and measure the F1 change.

Watch

Extended reading notes

Core claim

The central claim is that LLM pseudo labels, despite poor standalone accuracy, can be synergized with global label propagation to set a new state of the art in multimodal fake news detection. Concretely, GLPN-LLM (CLIP) raises the F1 score on Twitter to 89.03 against 85.97 for the naive FCN-LP (CLIP) + LLM baseline, with similar gains on PHEME and Weibo, and the ablation shows that the label-propagation module and the LLM module each contribute. The work frames the contribution as showing that the integration of LLMs is non-trivial but achievable through propagation rather than direct combination.

Load-bearing premise

The training-time random mask fully prevents label leakage, so at inference feeding the test node its own LLM pseudo label contributes propagated signal rather than an untrained shortcut.

Editorial extensions

If this is right

  • If the claim holds, LLM pseudo labels need not be accurate individually to help detection; propagation can amplify their collective signal.
  • The Global Random Mask with loss on masked nodes is a general recipe for label-integrated GCN training and should transfer to other semi-supervised graph tasks.
  • Confidence-filtered pseudo labels at the top 5% are enough; adding lower-confidence labels degrades performance.
  • The framework is practical: the LLM API cost per tweet is about $0.00074 and label propagation adds roughly 2% to computation time.

Reading between the lines

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

  • The reported gains could partly come from the test node's own pseudo-label embedding, which the model never saw in its own input during training; an ablation that zeroes test pseudo labels at inference would separate propagation from a shortcut.
  • The GRM's random masking and loss-on-masked-nodes rule leaves the weight path from a node's own non-zero label embedding to its prediction untrained, so the method's robustness to pseudo-label noise remains an open question.
  • A synthetic graph with known ground-truth labels could reveal whether propagation or the injected self-label drives the improvement; the paper's t-SNE evidence is suggestive but not quantitative.
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

2 major / 6 minor

Summary. The paper proposes GLPN-LLM, a framework that integrates LLM-generated pseudo-labels into a graph convolutional network via a global label propagation mechanism with a random masking strategy. The method augments each node's multimodal features with label-based features (ground-truth labels for training nodes and high-confidence LLM pseudo-labels for test nodes), then applies a GCN on a cross-modal similarity graph. The authors report state-of-the-art results on Twitter, PHEME, and Weibo, with the largest gains over the FCN-LP+LLM baseline. The central claim is that 'synergizing LLMs with label propagation' yields superior detection performance.

Significance. If the reported gains are genuine, the paper would offer a practical recipe for injecting LLM knowledge into graph-based fake-news detectors, with the additional strength of releasing code. The framework is clearly described and the experimental setup follows existing benchmarks. However, the significance is currently not established because the inference-time inclusion of a test node's own pseudo-label in its input features may create a shortcut that trivially reproduces the pseudo-label rather than performing label propagation. In addition, the two key hyperparameters are tuned on the same test sets used for final evaluation, so the reported numbers are optimistically biased. These issues must be resolved before the contribution can be assessed.

major comments (2)
  1. [Sections 3.4 and 3.5, Eqs. (2) and (5)] The inference procedure creates a self-loop shortcut that can manufacture the reported gains without label propagation. At inference, y'_i is set to \tilde y_i, and for high-confidence test nodes \tilde y_i is the node's own LLM pseudo-label (Eq. 5). This pseudo-label is concatenated into the node's input features (Eq. 2), and a standard GCN with self-loops aggregates a node's own features into its prediction. Because the loss is computed only on masked nodes during training, the model never learns to map a node's own label feature to its own class, but it does learn to map neighbor label features to a node's class; at test time the self-loop makes the node's own pseudo-label a neighbor feature, so the model can directly copy it. The global random mask (Eqs. 6-7) does not remove this test-time path. This threatens the central claim in the Abstract and Section 4.4, and the Limitation section does not acknowledge it. I request an experiment that blocks the self-pseudo-label path at inference (e.g., zeroing the label part of the self-loop or removing the self-loop for label features) and reports whether the improvements in Table 2 (Twitter F1 89.03 vs 86.30) persist; if they do not, the paper's conclusion is not supported.
  2. [Sections 4.6.1, 4.6.2, and Appendix C] The mask rate \lambda_mask and pseudo-label rate \lambda_pseudo are selected by sweeping on the same test sets that are used to produce the final results (Figure 3 and the values 0.50 and 0.05 in Appendix C). No held-out validation split is described. Consequently, the final numbers are tuned rather than independent predictions, and the comparison against baselines with fixed hyperparameters is unfair. The authors should select hyperparameters on a validation split (or via nested cross-validation) and then report test-set performance, and they should state the validation procedure explicitly in the experimental setup.
minor comments (6)
  1. [Section 3.5, Eqs. (6)-(7)] The statement that masking 'ensur[es] that training nodes do not propagate their own labels back to themselves' is imprecise: the mask zeroes the label only for the randomly selected subset, and the loss is computed only on those nodes. Unmasked nodes still carry their labels in their input features and propagate them to neighbors. Clarify the exact role of the mask and the loss computation.
  2. [Section 4.2 and experimental setup] The baseline 'FCN-LP + LLM' is described only as 'a naive solution, where FCN-LP directly uses LLM-generated pseudo labels.' The exact mechanism (e.g., whether pseudo-labels are used as node features, as post-processing, or as an ensemble) is not specified, making the comparison less reproducible.
  3. [Section 3.4, Eq. (3)] The LLM prompt in Eq. (3) uses only the cleaned text and does not include the image. Thus the LLM-generated pseudo-labels are text-only, despite the task being multimodal. This limitation should be stated explicitly, and the effect of using a multimodal LLM prompt could be discussed as future work.
  4. [Figure 3 and Section 4.6.2] The x-axis labels in the second row of Figure 3 mix proportions (0.01, 0.02, 0.03, 0.04, 0.05) and decimal values (0.1, 0.2, ..., 0.9). Clarify that these are proportions of the test set, and avoid the confusing use of the same numeric scale for both mask rate and pseudo-label rate.
  5. [Table 1 and Section 4.4] Several comparisons in Table 1 show overlapping standard deviations (e.g., PHEME F1 for GLPN-LLM (CLIP) 90.66±2.32 vs FCN-LP (CLIP)+LLM 89.21±2.59). The paper does not report significance tests or confidence intervals for the claimed improvements; adding such tests would strengthen the 'superior performance' claim.
  6. [Section 3.4] The term 'Mixed-Initiative Labeling' is not defined or standard in this context; consider renaming it to something more descriptive, such as 'Confidence-Based LLM Pseudo-Label Selection.'

Circularity Check

1 steps flagged · score 6.0 of 10

Test-time inclusion of each test node's own LLM pseudo-label in its GCN input lets reported gains be produced by copying that pseudo-label through the self-loop rather than by label propagation; the global random mask does not block this path at inference.

  1. self definitional [Section 3.4 Eqs. (2) and (5); Section 3.5 Global Random Mask; Section 4.6.2 (top-5% selection)]
    "During inference, we set the label-based features y′ i = ˜ yi. However, during training, y′ i is obtained via a Global Random Mask (GRM) to prevent label leakage. ... ˜yi = yi if nodei is truly labeled, ˆyi if unlabeled nodei is high-confidence, 0 otherwise. ... x′i = xi ⊕ y′i (2)"

    For a high-confidence test node i, Eq. (5) sets its label feature to its own LLM pseudo-label yhat_i; Section 3.5 makes y'_i = ytilde_i at inference; Eq. (2) concatenates y'_i into x'_i. A GCN aggregates a node's own feature through the self-loop, so the output logits for i are a direct learned function of yhat_i. The Global Random Mask (Eqs. 6-7) is applied only during training, so at inference this self-path is unmasked. Because only top-5% high-confidence pseudo-labels are injected (Sec. 4.6.2), those labels are accurate enough that copying them through the learned label-embedding submatrix can explain most of the reported improvement over FCN-LP+LLM (e.g., F1 89.03 vs 85.97 on Twitter).

full rationale

The central reported gains are partially circular in the specific sense that each high-confidence test node's own LLM pseudo-label is fed into that same node's GCN input at inference (Eqs. 2 and 5 vs. Section 3.5), while the masking mechanism that the paper credits for preventing label leakage is switched off exactly at inference. This is a concrete architectural reduction: the GCN's output for that node is a function of the label it is asked to predict, so the 'prediction' on the top-5% pseudo-labeled test nodes is not an independent result of label propagation. I score this as 6 rather than 8 or 10 because the method still propagates labels between nodes and content features for the remaining low-confidence nodes, so not every prediction is forced; the flaw is a partial, test-time leakage path rather than a fully definitional equivalence. I find no load-bearing self-citation or imported-uniqueness circularity: the graph construction and FCN-LP baseline are taken from external work (Zhao et al., 2023), and the core equations are the paper's own. Separately, Figure 3 selects lambda_mask and lambda_pseudo on the same test folds (reported results are averaged over the three benchmarks), which is an evaluation-leakage concern but not a circular-derivation issue and is not counted in the score. The Limitation section acknowledges dependence on backbone models and pseudo-label quality but does not disclose that test nodes' own pseudo-labels enter their own input features at inference, which is the circular path identified above.

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

The method rests on a few assumptions: the graph threshold inherited from FCN-LP, text-only LLM pseudo labels, and the key assumption that a training-time mask prevents inference-time self-label leakage. Two hyperparameters are selected on the test sets, which the paper reports honestly but which weakens the final numbers. No new entities are introduced.

free parameters (3)
  • mask rate lambda_mask = 0.50
    Selected by sweeping 0.1 to 0.9 on the test sets (Fig. 3, Sec. 4.6.1) and then used in the final model; this is test-set tuning that inflates the reported results.
  • pseudo label rate lambda_pseudo = 0.05
    Selected from a top 1% to 90% confidence sweep on the test sets (Fig. 3, Sec. 4.6.2) and used in the final model; another test-set tuned quantity.
  • similarity threshold theta = 0.95
    Inherited directly from FCN-LP (Zhao et al., 2023), not tuned in this paper, but it controls graph sparsity and is load-bearing for propagation.
assumptions (3)
  • ad hoc to paper Random masking during training is sufficient to prevent label leakage at inference, so including a test node's own pseudo label in its input features does not trivially determine its prediction.
    This is the load-bearing premise of the mask-based module; no experiment removes self-pseudo-labels at inference to verify it.
  • domain assumption Graph edges at cosine similarity threshold 0.95 connect semantically related news items and support label propagation.
    Inherited from FCN-LP; if the graph is too sparse or noisy, propagation gains disappear.
  • domain assumption LLM pseudo labels generated from text-only prompts are reliable enough to improve multimodal detection after filtering.
    The LLM uses only cleaned Twitter text in Eq. (3); image information is not used for pseudo labels.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Synergizing LLMs with Global Label Propagation for Multimodal Fake News Detection." pith.science (2026). https://pith.science/paper/L3JIPGW4

@misc{pith2026250600488,
  author       = {Pith},
  title        = {Pith review of: Synergizing LLMs with Global Label Propagation for Multimodal Fake News Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L3JIPGW4}},
  note         = {Machine review of arXiv:2506.00488}
}
read the original abstract

Large Language Models (LLMs) can assist multimodal fake news detection by predicting pseudo labels. However, LLM-generated pseudo labels alone demonstrate poor performance compared to traditional detection methods, making their effective integration non-trivial. In this paper, we propose Global Label Propagation Network with LLM-based Pseudo Labeling (GLPN-LLM) for multimodal fake news detection, which integrates LLM capabilities via label propagation techniques. The global label propagation can utilize LLM-generated pseudo labels, enhancing prediction accuracy by propagating label information among all samples. For label propagation, a mask-based mechanism is designed to prevent label leakage during training by ensuring that training nodes do not propagate their own labels back to themselves. Experimental results on benchmark datasets show that by synergizing LLMs with label propagation, our model achieves superior performance over state-of-the-art baselines.

Figures

Figures reproduced from arXiv: 2506.00488 by the authors.

Figure 1
Figure 1. Illustrations of different methods. et al., 2024; Bai et al., 2024), making them promis￾ing tools for enhancing fake news detection sys￾tems (Wu et al., 2024; Sun et al., 2024; Hu et al., 2024a; Su et al., 2023). A straightforward approach to leveraging LLMs for multimodal fake news detection involves di￾rectly combining predictions from existing models with LLM outputs, as illustrated in [PITH_FULL_IMAGE:figures/f… view at source ↗
Figure 2
Figure 2. Overview of the GLPN-LLM framework for fake news detection. The framework synergizes LLM [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The effect of the mask rate and the impact of the quantity of LLM-generated pseudo labels. The first [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Effect of prompt detail on performance. De [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Case study - examples of challenging in [PITH_FULL_IMAGE:figures/full_fig_p008_5.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. Maximizing Mutual Information Between Prompt and Response Improves LLM Performance With No Additional Data

    cs.LG 2026-03 unverdicted novelty 6.0 of 10

    MIPO constructs contrastive preference pairs from correct versus random prompts and uses DPO to maximize mutual information between prompts and responses, producing 3-40% gains on personalization and 1-18% on math tas...

Reference graph

Works this paper leans on

59 extracted references · 42 canonical work pages · cited by 1 Pith paper

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  2. [2]

    Jiaxin Bai, Yicheng Wang, Tianshi Zheng, Yue Guo, Xin Liu, and Yangqiu Song. 2024. Advancing abductive reasoning in knowledge graphs through complex logical hypothesis generation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024 , pages 1312-...

    work page 2024

  3. [3]

    Christina Boididou, Katerina Andreadou, Symeon Papadopoulos, Duc Tien Dang Nguyen, Giulia Boato, Michael Riegler, Yiannis Kompatsiaris, et al. 2015. Verifying multimedia use at mediaeval 2015. In MediaEval 2015, volume 1436. CEUR-WS

    work page 2015

  4. [4]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877--1901

    2020

  5. [5]

    Yuqi Chu, Lizi Liao, Zhiyuan Zhou, Chong-Wah Ngo, and Richang Hong. 2024. Towards multimodal emotional support conversation systems. arXiv preprint arXiv:2408.03650

    arXiv 2024

  6. [6]

    Quan Fang, Xiaowei Zhang, Jun Hu, Xian Wu, and Changsheng Xu. 2023. Contrastive multi-modal knowledge graph representation learning. IEEE Trans. Knowl. Data Eng. , 35(9):8983--8996

    work page 2023

  7. [7]

    Junyu Gao, Mengyuan Chen, and Changsheng Xu. 2025. Learning probabilistic presence-absence evidence for weakly-supervised audio-visual event perception. IEEE Transactions on Pattern Analysis and Machine Intelligence

    work page 2025

  8. [8]

    Beizhe Hu, Qiang Sheng, Juan Cao, Yuhui Shi, Yang Li, Danding Wang, and Peng Qi. 2024 a . Bad actor, good advisor: Exploring the role of large language models in fake news detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 22105--22113

    2024

Show all 59 references
  1. [9]

    Jun Hu, Bryan Hooi, and Bingsheng He. 2024 b . Efficient heterogeneous graph learning via random projection. IEEE Trans. Knowl. Data Eng. , 36(12):8093--8107

    2024

  2. [10]

    Jun Hu, Bryan Hooi, Bingsheng He, and Yinwei Wei. 2025. Modality-independent graph neural networks with global transformers for multimodal recommendation. In AAAI-25, Sponsored by the Association for the Advancement of Artificial Intelligence, February 25 - March 4, 2025, Phil...

    2025

  3. [11]

    Jun Hu, Bryan Hooi, Shengsheng Qian, Quan Fang, and Changsheng Xu. 2024 c . MGDCF: distance learning via markov graph diffusion for neural collaborative filtering. IEEE Trans. Knowl. Data Eng. , 36(7):3281--3296

    2024

  4. [12]

    Jun Hu, Shengsheng Qian, Quan Fang, Youze Wang, Quan Zhao, Huaiwen Zhang, and Changsheng Xu. 2021. Efficient graph deep learning in tensorflow with tf \_ geometric. In MM '21: ACM Multimedia Conference, Virtual Event, China, October 20 - 24, 2021 , pages 3775--3778. ACM

    2021

  5. [13]

    Qian Huang, Horace He, Abhay Singh, Ser - Nam Lim, and Austin R. Benson. 2021. Combining label propagation and simple models out-performs graph neural networks. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 . Open...

    2021

  6. [14]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276

    2024 arXiv

  7. [15]

    Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. 2019. Label propagation for deep semi-supervised learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5070--5079

    2019

  8. [16]

    Zhiwei Jin, Juan Cao, Han Guo, Yongdong Zhang, and Jiebo Luo. 2017. Multimodal fusion with recurrent neural networks for rumor detection on microblogs. In Proceedings of the 25th ACM international conference on Multimedia, pages 795--816

    2017

  9. [17]

    Dhruv Khattar, Jaipal Singh Goud, Manish Gupta, and Vasudeva Varma. 2019. Mvae: Multimodal variational autoencoder for fake news detection. In The world wide web conference, pages 2915--2921

    2019

  10. [18]

    Diederik P Kingma. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

    2014 arXiv

  11. [19]

    Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907

    2016 arXiv

  12. [20]

    Yuxuan Liang, Wentao Zhang, Zeang Sheng, Ling Yang, Jiawei Jiang, Yunhai Tong, and Bin Cui. 2024. HGAMLP: heterogeneous graph attention MLP with de-redundancy mechanism. In 40th IEEE International Conference on Data Engineering, ICDE 2024, Utrecht, The Netherlands, May 13-16, ...

    2024

  13. [21]

    Qiheng Mao, Zemin Liu, Chenghao Liu, and Jianling Sun. 2023. Hinormer: Representation learning on heterogeneous information networks with graph transformer. In Proceedings of the ACM Web Conference 2023, WWW 2023, Austin, TX, USA, 30 April 2023 - 4 May 2023 , pages 599--610. ACM

    2023

  14. [22]

    Ver \'o nica P \'e rez-Rosas, Bennett Kleinberg, Alexandra Lefevre, and Rada Mihalcea. 2017. Automatic detection of fake news. arXiv preprint arXiv:1708.07104

    2017 arXiv

  15. [23]

    Shengsheng Qian, Jinguang Wang, Jun Hu, Quan Fang, and Changsheng Xu. 2021. Hierarchical multi-modal contextual attention network for fake news detection. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval, pages 153--162

    2021

  16. [24]

    Hezhe Qiao, Hanghang Tong, Bo An, Irwin King, Charu Aggarwal, and Guansong Pang. 2024. Deep graph anomaly detection: A survey and new perspectives. arXiv preprint arXiv:2409.09957

    2024 arXiv

  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]

    Lei Sang, Honghao Li, Yiwen Zhang, Yi Zhang, and Yun Yang. 2025 a . Adagin: Adaptive graph interaction network for click-through rate prediction. ACM Trans. Inf. Syst. , 43(1):3:1--3:31

    2025

  19. [27]

    Lei Sang, Yu Wang, Yi Zhang, Yiwen Zhang, and Xindong Wu. 2025 b . Intent-guided heterogeneous graph contrastive learning for recommendation. IEEE Trans. Knowl. Data Eng. , 37(4):1915--1929

    2025

  20. [28]

    Kai Shu, Amy Sliva, Suhang Wang, Jiliang Tang, and Huan Liu. 2017. Fake news detection on social media: A data mining perspective. ACM SIGKDD Explorations Newsletter, 19(1):22--36

    2017

  21. [29]

    Kai Shu, Suhang Wang, and Huan Liu. 2019. Beyond news contents: The role of social context for fake news detection. In Proceedings of the Twelfth ACM International Conference on Web Search and Data Mining, pages 312--320. ACM

    2019

  22. [30]

    Shivangi Singhal, Rajiv Ratn Shah, Tanmoy Chakraborty, Ponnurangam Kumaraguru, and Shin'ichi Satoh. 2019. Spotfake: A multi-modal framework for fake news detection. In 2019 IEEE fifth international conference on multimedia big data (BigMM), pages 39--47. IEEE

    2019

  23. [31]

    Jinyan Su, Claire Cardie, and Preslav Nakov. 2023. Adapting fake news detection to the era of large language models. arXiv preprint arXiv:2311.04917

    2023 arXiv

  24. [32]

    Chuxiong Sun, Jie Hu, Hongming Gu, Jinpeng Chen, Wei Liang, and Mingchuan Yang. 2025. Scalable and adaptive graph neural networks with self-label-enhanced training. Pattern Recognit., 160:111210

    2025

  25. [33]

    Peijie Sun, Le Wu, Kun Zhang, Xiangzhi Chen, and Meng Wang. 2023. Neighborhood-enhanced supervised contrastive learning for collaborative filtering. IEEE Transactions on Knowledge and Data Engineering, 36(5):2069--2081

    2023

  26. [34]

    Yanshen Sun, Jianfeng He, Limeng Cui, Shuo Lei, and Chang-Tien Lu. 2024. Exploring the deceptive power of llm-generated fake news: A study of real-world detection challenges. arXiv preprint arXiv:2403.18249

    2024 arXiv

  27. [35]

    Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li \` o , and Yoshua Bengio. 2018. Graph attention networks. In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Pro...

    2018

  28. [36]

    Yaqing Wang, Fenglong Ma, Zhiwei Jin, Ye Yuan, Guangxu Xun, Kishlay Jha, Lu Su, and Jing Gao. 2018. Eann: Event adversarial neural networks for multi-modal fake news detection. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining,...

    2018

  29. [37]

    Youze Wang, Shengsheng Qian, Jun Hu, Quan Fang, and Changsheng Xu. 2020. Fake news detection via knowledge-driven multimodal graph convolutional networks. In Proceedings of the 2020 on International Conference on Multimedia Retrieval, ICMR 2020, Dublin, Ireland, June 8-11, 202...

    2020

  30. [38]

    Jiaying Wu, Jiafeng Guo, and Bryan Hooi. 2024. Fake news in sheep's clothing: Robust fake news detection against llm-empowered style attacks. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3367--3378

    2024

  31. [39]

    Yang Wu, Pengwei Zhan, Yunjian Zhang, Liming Wang, and Zhen Xu. 2021. Multimodal fusion with co-attention networks for fake news detection. In Findings of the association for computational linguistics: ACL-IJCNLP 2021, pages 2560--2569

    2021

  32. [40]

    Siheng Xiong, Ali Payani, Ramana Kompella, and Faramarz Fekri. 2024. Large language models can learn temporal reasoning. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-1...

    2024

  33. [41]

    Yifan Xu, Xiaoshan Yang, Yaguang Song, and Changsheng Xu. 2024. Libra: Building decoupled vision system on large language models. In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024 . OpenReview.net

    2024

  34. [42]

    Xiaocheng Yang, Mingyu Yan, Shirui Pan, Xiaochun Ye, and Dongrui Fan. 2023. Simple and efficient heterogeneous graph neural network. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications of Artificial Intell...

    2023

  35. [43]

    Xiaoshan Yang, Baochen Xiong, Yi Huang, and Changsheng Xu. 2024. Cross-modal federated human activity recognition. IEEE Trans. Pattern Anal. Mach. Intell. , 46(8):5345--5361

    2024

  36. [44]

    Xiaoyu Yang, Yuefei Lyu, Tian Tian, Yifei Liu, Yudong Liu, and Xi Zhang. 2021. Rumor detection on social media with graph structured adversarial learning. In Proceedings of the twenty-ninth international conference on international joint conferences on artificial intelligence,...

    2021

  37. [45]

    Huaiwen Zhang, Zihang Guo, Yang Yang, Xin Liu, and De Hu. 2023. C\( ^ 2 \)st: Cross-modal contextualized sequence transduction for continuous sign language recognition. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023 , pages...

    2023

  38. [46]

    Huaiwen Zhang, Xinxin Liu, Qing Yang, Yang Yang, Fan Qi, Shengsheng Qian, and Changsheng Xu. 2024. T\( ^ 3 \)rd: Test-time training for rumor detection on social media. In Proceedings of the ACM on Web Conference 2024, WWW 2024, Singapore, May 13-17, 2024 , pages 2407--2416. ACM

    2024

  39. [47]

    Wentao Zhang, Ziqi Yin, Zeang Sheng, Yang Li, Wen Ouyang, Xiaosen Li, Yangyu Tao, Zhi Yang, and Bin Cui. 2022. Graph attention multi-layer perceptron. In KDD '22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, August 14 - 18, 2022 ,...

    2022

  40. [48]

    Xichen Zhang and Ali A Ghorbani. 2020. An overview of online fake news: Characterization, detection, and discussion. Information Processing & Management, 57(2):102025

    2020

  41. [49]

    Zhen Zhang and Bingsheng He. 2025. Aggregate to adapt: Node-centric aggregation for multi-source-free graph domain adaptation. In Proceedings of the ACM on Web Conference 2025, WWW 2025, Sydney, NSW, Australia, 28 April 2025- 2 May 2025 , pages 4420--4431. ACM

    2025

  42. [50]

    Wanqing Zhao, Yuta Nakashima, Haiyuan Chen, and Noboru Babaguchi. 2023. Enhancing fake news detection in social media via label propagation on cross-modal tweet graph. In Proceedings of the 31st ACM International Conference on Multimedia, pages 2400--2408

    2023

  43. [51]

    Dengyong Zhou, Olivier Bousquet, Thomas Lal, Jason Weston, and Bernhard Sch \"o lkopf. 2003. Learning with local and global consistency. Advances in neural information processing systems, 16

    2003

  44. [52]

    Xinyi Zhou, Jindi Wu, and Reza Zafarani. 2020. : Similarity-aware multi-modal fake news detection. In Pacific-Asia Conference on knowledge discovery and data mining, pages 354--367. Springer

    2020

  45. [53]

    Xinyi Zhou and Reza Zafarani. 2020. A survey of fake news: Fundamental theories, detection methods, and opportunities. ACM Computing Surveys (CSUR), 53(5):1--40

    2020

  46. [54]

    Xinyi Zhou, Reza Zafarani, Kai Shu, and Huan Liu. 2019. Fake news: Fundamental theories, detection strategies and challenges. In Proceedings of the twelfth ACM international conference on web search and data mining, pages 836--837

    2019

  47. [55]

    Xiaojin Zhu and Zoubin Ghahramani. 2002. Learning from labeled and unlabeled data with label propagation. ProQuest number: information to all users

    2002

  48. [56]

    Xiaojin Zhu, Zoubin Ghahramani, and John D Lafferty. 2003. Semi-supervised learning using gaussian fields and harmonic functions. In Proceedings of the 20th International conference on Machine learning (ICML-03), pages 912--919

    2003

  49. [57]

    Arkaitz Zubiaga, Maria Liakata, and Rob Procter. 2017. Exploiting context for rumour detection in social media. In Social Informatics: 9th International Conference, SocInfo 2017, Oxford, UK, September 13-15, 2017, Proceedings, Part I 9, pages 109--123. Springer

    2017

  50. [58]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  51. [59]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.