Pith. sign in

REVIEW 4 major objections 3 minor 42 references

LLM-based Contrastive Self-Supervised AMR Learning with Masked Graph Autoencoders for Fake News Detection

T0 review · 4 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims a self-supervised system can detect fake news by learning from AMR semantic graphs plus social propagation graphs, without supervised training of the representations.

desk verdict The architecture is genuinely interesting, but the empirical core is not reliable as presented — the paper's own numbers contradict each other on the split protocol, so the SOTA claim is currently unsupported. read the letter →

arxiv 2508.18819 v1 pith:GUPNVQSY submitted 2025-08-26 cs.CL cs.SI

Shubham Gupta , Shraban Kumar Chatterjee , Suman Kundu This is my paper
classification cs.CLcs.SI
keywords fakenewsdetectionself-supervisedlearningAbstractMeaningRepresentation(AMR)graphcontrastivemaskedautoencoderLLMpseudo-labelssocialpropagationNet
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that fake news can be detected without supervised pretraining of the feature extractors, by combining two self-supervised views of an article: an Abstract Meaning Representation (AMR) graph of its meaning, and the social graph of how it spread. The core move is an LLM-based graph contrastive loss (LGCL): an LLM guesses which articles are real or fake in zero-shot fashion, those guesses are used to build class centroids of fixed BERT features, and the AMR graph encoder is pulled toward the article's BERT feature while pushed away from the opposite class's centroid. A multi-view graph masked autoencoder learns propagation features from augmented social graphs. On FakeNewsNet, the paper reports accuracy/F1 of 0.919/0.918 on PolitiFact and 0.968/0.966 on GossipCop, exceeding the previous unsupervised benchmark GAMC and, on these datasets, the supervised EA2N-BERT baseline. If the claim holds, semantic structure plus propagation dynamics learned without labels rival or beat supervised training for fake news detection, especially when labeled data are scarce.

What carries the argument

The central mechanism is the LLM-based graph contrastive loss (LGCL), built on the Scaled Cosine Error (SCE). A graph transformer encodes the WikiAMR graph; SCE pulls the encoded embedding toward the fixed BERT feature of the same text (the positive target), while a margin term pushes it away from an LLM-selected negative centroid, namely the mean BERT feature of the opposite class as judged by the LLM's zero-shot pseudo-labels. The propagation branch is a multi-view graph masked autoencoder: two augmented views of the social graph are encoded by a GIN, remasked multiple times, and decoded, with a reconstruction loss plus a cross-view cosine-similarity loss. Concatenating the semantic and pr

What would settle it

Run the full pipeline with the LLM pseudo-labels shuffled at random, keeping all other hyperparameters fixed. If PolitiFact accuracy stays near 0.919 and GossipCop near 0.968, the LLM-based contrastive loss is not the source of the gain; if accuracy drops markedly, the contrastive anchors are load-bearing. A complementary check is replacing the LLM centroids with oracle true-label centroids and measuring the difference.

Watch

Extended reading notes

Core claim

The central claim: a self-supervised model detects fake news by jointly learning an AMR-based semantic structure and a propagation-graph structure. A graph transformer encodes the WikiAMR graph; the LLM-based graph contrastive loss (LGCL) pulls this embedding toward the article's fixed BERT feature and pushes it away from an LLM-selected negative centroid (fake centroid for real articles, real centroid for fake articles). A multi-view masked graph autoencoder learns propagation features from augmented social graphs; the concatenated features feed a linear SVM. Reported accuracy/F1: 0.919/0.918 PolitiFact, 0.968/0.966 GossipCop, beating GAMC (0.838/0.831, 0.946/0.943) and supervised EA2N-BERT

Load-bearing premise

The AMR branch assumes that the LLM's zero-shot real/fake labels, which the paper reports as only about 80% and 68% accurate, still produce class centroids of fixed BERT features whose opposite-class direction is a useful negative direction in feature space.

Editorial extensions

If this is right

  • Labeled-data cost drops: with only 10% of labeled training data, the method still reaches 0.875 accuracy on PolitiFact and 0.951 on GossipCop, so low-resource deployment is plausible.
  • The full model beats the prior unsupervised benchmark GAMC by 8.1 points of accuracy on PolitiFact and 2.2 points on GossipCop.
  • It also beats the supervised EA2N-BERT baseline on both benchmarks (0.919 vs 0.911 on PolitiFact; 0.968 vs 0.844 on GossipCop).
  • Ablations show both branches contribute: LGCL alone gives 0.841/0.948 and propagation alone gives 0.846/0.946, while the combined model reaches 0.919/0.968.
  • The authors argue the same self-supervised graph recipe transfers to question answering, event detection, hate speech, and aggression detection.

Reading between the lines

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

  • The propagation branch alone already reaches 0.846/0.945, close to GAMC, so on GossipCop much of the signal may come from social dynamics; the AMR branch likely matters most when the article text itself is the decisive cue.
  • The reported silhouette scores (0.16 to 0.40 on GossipCop, 0.33 to 0.64 on PolitiFact) suggest the mechanism is geometric feature separation, which makes the pipeline a useful probe for how much of the gain comes from the margin term versus the reconstruction term.
  • A direct stress test would be to shuffle the LLM's pseudo-labels (or artificially raise their error rate) and watch how quickly the LGCL margin term stops helping; this would quantify the method's tolerance for noisy anchors.
  • Because the LLM only needs to provide a weak binary partition, the negative-sampler recipe could be reused for any scarce-label binary problem where a cheap zero-shot model can seed contrastive anchors.
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 / 3 minor

Summary. The paper proposes a self-supervised fake news detection framework that combines semantic features from Abstract Meaning Representation (AMR) graphs with propagation features from social context graphs. The AMR branch is trained with a novel LLM-based graph contrastive loss (LGCL) that uses LLaMA's zero-shot pseudo-labels to compute negative centroids, while the propagation branch uses a multi-view masked graph autoencoder. The two sets of features are concatenated and classified by a linear SVM. Experiments on PolitiFact and GossipCop report state-of-the-art accuracy/F1 (e.g., 0.919/0.918 on PolitiFact, 0.968/0.966 on GossipCop), outperforming both unsupervised baselines such as GAMC and supervised methods such as EA2N-BERT, with ablations and t-SNE visualizations.

Significance. The architectural combination is novel: integrating AMR semantic graphs with a multi-view graph autoencoder and an LLM-driven contrastive sampler is a reasonable research direction for low-resource fake news detection. The paper includes several strengths: (i) it explicitly targets the label-scarcity setting; (ii) it provides ablations for the main components (Table 6); (iii) it reports variance for the proposed method; and (iv) it gives qualitative t-SNE and silhouette scores showing progressive feature separation. If the empirical claims are reproduced with a clean and consistent protocol, the work would be a useful contribution to graph-based misinformation detection. However, the current evidence is undermined by inconsistencies in the reported experimental protocol and by the absence of error bars for baselines.

major comments (4)
  1. [Table 5 vs. Tables 2–4; Implementation Details] The headline results are not reproducible from the reported protocol. Tables 2 and 4 list PolitiFact Acc=0.919, F1=0.918 and GossipCop Acc=0.968, F1=0.966 as the main results. Table 5, at 'Train Size % = 80', reports PolitiFact Acc=0.938, F1=0.938 and GossipCop Acc=0.954, F1=0.955. The Implementation Details say both 'our main results are based on an 80:20 train-test split' and 'reported the results from 80% of the training data with 5-fold cross-validation', while the text around Table 5 says 'keeping the test set fixed at 10%'. These statements are mutually incompatible, and no row of Table 5 matches the headline numbers. The paper must specify exactly which split produced Tables 2–4 and reconcile Table 5, or the central claim of SOTA performance is not verifiable.
  2. [Figures 2–3 and hyperparameter selection] The hyperparameters λ (Eq. 4), k, and m are selected based on accuracy plots in Figures 2 and 3. The text states that accuracy is maximized at λ=0.5 and k=2, m=2. No validation split is described; if these values were chosen by examining test-set accuracy, all reported numbers are optimistically biased. The authors should describe a proper validation procedure (e.g., a held-out validation subset or nested cross-validation) and report results for hyperparameters chosen on validation data only.
  3. [Tables 2–3: missing baseline variance] The variance rows in Tables 2 and 3 report ± values only for the proposed method. Baselines (GAMC, GTUT, etc.) are quoted as single point estimates without standard deviations or number of seeds. Given that the claimed gains over GAMC are 8.1% accuracy on PolitiFact, a statistically grounded comparison is essential. Please provide mean±std over multiple runs for all methods, or at least conduct a paired significance test against the strongest baselines.
  4. [Table 6 and LLM-based Negative Sampler] The contrastive loss in Eq. (4) relies on negative centroids computed from LLaMA's zero-shot pseudo-labels. Table 6 reports that LLaMA's zero-shot accuracy is only 0.804 on PolitiFact and 0.680 on GossipCop. If the pseudo-labels are systematically biased (e.g., consistently misclassifying celebrity gossip as real), the centroids would not separate real and fake in the intended direction, and LGCL could push embeddings together rather than apart. The paper does not analyze the sensitivity of the method to pseudo-label noise. A robustness experiment (e.g., randomly flipping a fraction of pseudo-labels, or comparing against random centroids) is needed to support the claim that the LLM-based sampling is beneficial.
minor comments (3)
  1. [Text near Table 4] The text states 'our model achieves an accuracy of 0.919 and an F1-score of 0.933' on PolitiFact, but Table 4 reports F1=0.918. Please correct the discrepancy.
  2. [Implementation Details] The phrase 'reported the results from 80 % of the training data with 5-fold cross-validation' is confusing. Please clarify whether the final SVM uses 5-fold cross-validation on 80% of the data as training, or an 80:20 split with a single test set.
  3. [Equation 8 and feature combination] The multi-view cosine similarity loss in Eq. (8) has unwieldy notation and the domain of the mean is not fully specified. Also, the final feature combination is written as 'HGamr · HGprop' in the text but as 'concatenating' in Figure 1 and elsewhere; please use a consistent symbol (e.g., concatenation ⊕) and define it.

Circularity Check

0 steps flagged · score 0.0 of 10

No derivation step reduces to its inputs by construction; minor self-citations are not load-bearing and the reported empirical concerns are reproducibility issues, not circularity.

full rationale

The AMR branch trains a graph transformer with LGCL to reconstruct the fixed BERT feature ypos and to repel an LLM-computed centroid yneg; the propagation branch uses a masked graph autoencoder to reconstruct the input node features. In both cases the training signal is an input feature or a pseudo-label centroid, not the ground-truth label used in the final SVM, so the reported accuracy is not identical to the LLM zero-shot output or to the BERT input. The self-citations to (Gupta, Rajora, and Kundu 2025) supply the Evidence Linking Algorithm and its hyperparameters from the authors' own published, externally evaluated system; the current paper reuses this component rather than invoking an unverified uniqueness claim or smuggling in an ansatz, so real evidence supports the citation and the score is not raised. No equation is identical by construction to another equation or to the fitted pseudo-labels. The manuscript does contain important non-circular weaknesses: Table 5's 80% train row (PolitiFact Acc=0.938, F1=0.938) disagrees with the headline Tables 2/4 (Acc=0.919, F1=0.918), and hyperparameters λ, k, m are selected using accuracy curves with no described validation split, which may bias the reported test numbers. These are reproducibility and validity concerns, not circularity of the derivation chain.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The framework's central claim rests on the transfer of several pretrained components: an AMR parser, BERT embeddings, an LLM, and the authors' own evidence linking module. The free parameters (lambda, margin, gamma, k, m, augmentation ratios) were set without a separate validation set, and the paper does not report a sweep protocol. No new entity is postulated.

free parameters (7)
  • lambda (contrastive weight); also written alpha in text = 0.5
    Weight for negative term in Eq. 4; selected by maximizing test accuracy (Figure 2).
  • margin m in Eq. 4
    Margin for pushing negatives apart in cosine space; value never reported.
  • gamma in SCE (Eq. 3)
    Scaling factor in scaled cosine error, taken from GraphMAE; value not stated.
  • k (number of graph augmentations) = 2
    Chosen from PolitiFact accuracy grid (Figure 3).
  • m (number of multi-view remaskings) = 2
    Chosen from PolitiFact accuracy grid (Figure 3).
  • augmentation ratios (50% masking, 20% edge drop)
    Graph augmentation hyperparameters in propagation module.
  • GIN layers (2 encoder, 1 decoder)
    Architecture depth for the propagation autoencoder.
assumptions (6)
  • domain assumption The STOG AMR parser produces meaning graphs whose relations are useful for veracity classification.
    Invoked at the start of Self-supervised AMR Graph Learning; if AMR graphs from noisy social text are poor, the AMR branch is moot.
  • domain assumption The Evidence Linking Algorithm from EA2N (Gupta, Rajora, Kundu 2025) creates WikiAMR graphs that help distinguish real from fake, and its parameters transfer to the self-supervised setting.
    Section 'Self-supervised AMR Graph Learning' reuses ELA with 'the same parameters'; this is a black-box borrow from the authors' supervised paper.
  • domain assumption LLaMA-3-7B zero-shot pseudo-labels are informative enough that BERT-feature centroids per pseudo-class form meaningful negative anchors.
    Section 'LLM-based Negative Sampler'; the paper's Table 6 shows LLaMA zero-shot accuracy 0.68 on GossipCop, so this assumption is non-trivial.
  • domain assumption Fixed BERT text features are a good positive target for the AMR graph encoder.
    In Eq. 4, y_pos is the original BERT-derived feature; if text and AMR geometry are misaligned, training pulls the graph encoder toward an arbitrary space.
  • domain assumption Propagation graph node features (news BERT, user recent-200-posts) and the masked autoencoder objective capture veracity-relevant social dynamics.
    Section 'Multi-View Social Context and Propagation Graph Learning'; borrowed from GAMC/Dou et al.
  • domain assumption FakeNewsNet ground-truth labels are reliable enough to judge the method.
    Experiments use PolitiFact and GossipCop; the paper itself notes 'datasets have known issues with label reliability', which cuts both ways for the evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLM-based Contrastive Self-Supervised AMR Learning with Masked Graph Autoencoders for Fake News Detection." pith.science (2026). https://pith.science/paper/GUPNVQSY

@misc{pith2026250818819,
  author       = {Pith},
  title        = {Pith review of: LLM-based Contrastive Self-Supervised AMR Learning with Masked Graph Autoencoders for Fake News Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GUPNVQSY}},
  note         = {Machine review of arXiv:2508.18819}
}
read the original abstract

The proliferation of misinformation in the digital age has led to significant societal challenges. Existing approaches often struggle with capturing long-range dependencies, complex semantic relations, and the social dynamics influencing news dissemination. Furthermore, these methods require extensive labelled datasets, making their deployment resource-intensive. In this study, we propose a novel self-supervised misinformation detection framework that integrates both complex semantic relations using Abstract Meaning Representation (AMR) and news propagation dynamics. We introduce an LLM-based graph contrastive loss (LGCL) that utilizes negative anchor points generated by a Large Language Model (LLM) to enhance feature separability in a zero-shot manner. To incorporate social context, we employ a multi view graph masked autoencoder, which learns news propagation features from social context graph. By combining these semantic and propagation-based features, our approach effectively differentiates between fake and real news in a self-supervised manner. Extensive experiments demonstrate that our self-supervised framework achieves superior performance compared to other state-of-the-art methodologies, even with limited labelled datasets while improving generalizability.

Figures

Figures reproduced from arXiv: 2508.18819 by the authors.

Figure 1
Figure 1. Overview of the proposed method: The news article is converted to an AMR graph [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Change in classification result with different values [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. The TSNE plots showing the embeddings of Poli [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 38 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Banarescu, L.; Bonial, C.; Cai, S.; Georgescu, M.; Griffitt, K.; Hermjakob, U.; Knight, K.; Koehn, P.; Palmer, M.; and Schneider, N. 2013. A bstract M eaning R epresentation for Sembanking. In Proceedings of the 7th Linguistic Annotation Workshop and Interoperability with Discourse, 178--186. Sofia, Bulgaria

    work page 2013

  4. [4]

    Cai, D.; and Lam, W. 2020. Graph Transformer for Graph-to-Sequence Learning. In AAAI, 7464--7471. AAAI Press

    work page 2020

  5. [5]

    Cho, K.; van Merri \"e nboer, B.; Gulcehre, C.; Bahdanau, D.; Bougares, F.; Schwenk, H.; and Bengio, Y. 2014. Learning Phrase Representations using RNN Encoder -- Decoder for Statistical Machine Translation. In EMNLP, 1724--1734. Doha, Qatar: ACL

    work page 2014

  6. [6]

    Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2019. BERT : Pre-training of Deep Bidirectional Transformers for Language Understanding. In Burstein, J.; Doran, C.; and Solorio, T., eds., Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and ...

    work page 2019

  7. [7]

    Dewatana, H.; and Adillah, S. U. 2021. The effectiveness of criminal eradication on hoax information and fake news. Law Development Journal, 3(3): 513--520

    work page 2021

  8. [8]

    S.; and Sun, L

    Dou, Y.; Shu, K.; Xia, C.; Yu, P. S.; and Sun, L. 2021. User Preference-aware Fake News Detection. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '21, 2051–2055. New York, NY, USA: Association for Computing Machinery. ISBN 9781450380379

    work page 2021

Show all 42 references
  1. [9]

    Dun, Y.; Tu, K.; Chen, C.; Hou, C.; and Yuan, X. 2021. KAN: Knowledge-aware Attention Network for Fake News Detection. AAAI, 35(1): 81--89

    2021

  2. [10]

    Feng, S.; Banerjee, R.; and Choi, Y. 2012. Syntactic Stylometry for Deception Detection. In ACL (Volume 2: Short Papers), 171--175. Jeju Island, Korea: ACL

    2012

  3. [11]

    Gangireddy, S. C. R.; P, D.; Long, C.; and Chakraborty, T. 2020. Unsupervised Fake News Detection: A Graph-based Approach. In Proceedings of the 31st ACM Conference on Hypertext and Social Media, HT '20, 75–83. New York, NY, USA: Association for Computing Machinery. ISBN 9781450370981

    2020

  4. [12]

    P.; Rosso, P.; and Rangel, F

    Ghanem, B.; Ponzetto, S. P.; Rosso, P.; and Rangel, F. 2021. FakeFlow: Fake News Detection by Modeling the Flow of Affective Information. In 16th EACL

    2021

  5. [13]

    Gupta, S.; Rajora, A.; and Kundu, S. 2025. EA2N: Evidence-based AMR Attention Network for Fake News Detection. IEEE Transactions on Knowledge and Data Engineering, 1--12

    2025

  6. [14]

    Gupta, S.; Yadav, N.; Kundu, S.; and Sankepally, S. 2023. FakEDAMR: Fake News Detection Using Abstract Meaning Representation Network. In International Conference on Complex Networks and Their Applications, 308--319. Springer

    2023

  7. [15]

    Hou, Z.; He, Y.; Cen, Y.; Liu, X.; Dong, Y.; Kharlamov, E.; and Tang, J. 2023. GraphMAE2: A Decoding-Enhanced Masked Self-Supervised Graph Learner. In Proceedings of the ACM Web Conference 2023, WWW '23, 737–746. New York, NY, USA: Association for Computing Machinery. ISBN 978...

    2023

  8. [16]

    Hou, Z.; Liu, X.; Cen, Y.; Dong, Y.; Yang, H.; Wang, C.; and Tang, J. 2022. GraphMAE: Self-Supervised Masked Graph Autoencoders. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD '22, 594–604. New York, NY, USA: Association for Comput...

    2022

  9. [17]

    Hu, L.; Yang, T.; Zhang, L.; Zhong, W.; Tang, D.; Shi, C.; Duan, N.; and Zhou, M. 2021. Compare to The Knowledge: Graph Neural Fake News Detection with External Knowledge. In ACL-IJCNLP (Volume 1: Long Papers), 754--763. Online: ACL

    2021

  10. [18]

    N.; and Welling, M

    Kipf, T. N.; and Welling, M. 2016. Variational Graph Auto-Encoders. arXiv:1611.07308

    2016 arXiv

  11. [19]

    Li, D.; Guo, H.; Wang, Z.; and Zheng, Z. 2021. Unsupervised Fake News Detection Based on Autoencoder. IEEE Access, 9: 29356--29365

    2021

  12. [20]

    M.; and Tong, W

    Li, S.; Li, W.; Luvembe, A. M.; and Tong, W. 2024. Graph Contrastive Learning With Feature Augmentation for Rumor Detection. IEEE Transactions on Computational Social Systems, 11(4): 5158--5167

    2024

  13. [21]

    Li, Z.; Liu, F.; Yang, W.; Peng, S.; and Zhou, J. 2022. A Survey of Convolutional Neural Networks: Analysis, Applications, and Prospects. IEEE Transactions on Neural Networks and Learning Systems, 33(12): 6999--7019

    2022

  14. [22]

    Liu, Y.; and Wu, Y.-F. 2018. Early Detection of Fake News on Social Media Through Propagation Path Classification with Recurrent and Convolutional Networks. AAAI, 32(1)

    2018

  15. [23]

    Long, Y.; Lu, Q.; Xiang, R.; Li, M.; and Huang, C.-R. 2017. Fake News Detection Through Multi-Perspective Speaker Profiles. In IJCNLP (Volume 2: Short Papers), 252--256. Taipei, Taiwan: Asian Federation of Natural Language Processing

    2017

  16. [24]

    J.; Wong, K.-F.; and Cha, M

    Ma, J.; Gao, W.; Mitra, P.; Kwon, S.; Jansen, B. J.; Wong, K.-F.; and Cha, M. 2016 a . Detecting rumors from microblogs with recurrent neural networks. In Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence, IJCAI'16, 3818–3824. AAAI Press...

    2016

  17. [25]

    J.; Wong, K.-F.; and Cha, M

    Ma, J.; Gao, W.; Mitra, P.; Kwon, S.; Jansen, B. J.; Wong, K.-F.; and Cha, M. 2016 b . Detecting Rumors from Microblogs with Recurrent Neural Networks. In IJCAI, IJCAI'16, 3818–3824. AAAI Press. ISBN 9781577357704

    2016

  18. [26]

    Min, E.; Rong, Y.; Bian, Y.; Xu, T.; Zhao, P.; Huang, J.; and Ananiadou, S. 2022. Divide-and-Conquer: Post-User Interaction Network for Fake News Detection on Social Media. In Proceedings of the ACM Web Conference 2022, WWW '22, 1148–1158. New York, NY, USA: Association for Co...

    2022

  19. [27]

    Popat, K.; Mukherjee, S.; Str\" o tgen, J.; and Weikum, G. 2017. Where the Truth Lies: Explaining the Credibility of Emerging Claims on the Web and Social Media. WWW '17 Companion, 1003–1012. Republic and Canton of Geneva, CHE: International World Wide Web Conferences Steering...

    2017

  20. [28]

    Shu, K.; Cui, L.; Wang, S.; Lee, D.; and Liu, H. 2019. dEFEND: Explainable Fake News Detection. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD '19, 395–405. New York, NY, USA: Association for Computing Machinery. ISBN 9...

    2019

  21. [29]

    Shu, K.; Mahudeswaran, D.; Wang, S.; Lee, D.; and Liu, H. 2020. FakeNewsNet: A Data Repository with News Content, Social Context, and Spatiotemporal Information for Studying Fake News on Social Media. Big Data, 8(3): 171--188

    2020

  22. [30]

    Silva, A.; Luo, L.; Karunasekera, S.; and Leckie, C. 2024. Unsupervised Domain-Agnostic Fake News Detection Using Multi-Modal Weak Signals . IEEE Transactions on Knowledge & Data Engineering, 36(11): 7283--7295

    2024

  23. [31]

    Sun, T.; Qian, Z.; Dong, S.; Li, P.; and Zhu, Q. 2022. Rumor Detection on Social Media with Graph Adversarial Contrastive Learning. In Proceedings of the ACM Web Conference 2022, WWW '22, 2789–2797. New York, NY, USA: Association for Computing Machinery. ISBN 9781450390965

    2022

  24. [32]

    Wang, Y.; Ma, F.; Jin, Z.; Yuan, Y.; Xun, G.; Jha, K.; Su, L.; and Gao, J. 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, KDD '18, 849–857. ...

    2018

  25. [33]

    Wu, L.; Lin, H.; Tan, C.; Gao, Z.; and Li, S. Z. 2023. Self-Supervised Learning on Graphs: Contrastive, Generative, or Predictive. 35(4): 4216–4235

    2023

  26. [34]

    Xu, K.; Hu, W.; Leskovec, J.; and Jegelka, S. 2019. How Powerful are Graph Neural Networks? In International Conference on Learning Representations

    2019

  27. [35]

    Yang, R.; Wang, X.; Jin, Y.; Li, C.; Lian, J.; and Xie, X. 2022. Reinforcement Subgraph Reasoning for Fake News Detection. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD '22, 2253–2262. New York, NY, USA: Association for Computing ...

    2022

  28. [36]

    Yao, H.; Wu, Y.-x.; Al-Shedivat, M.; and Xing, E. 2021. Knowledge-Aware Meta-learning for Low-Resource Text Classification. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 1814--1821. Online and Punta Cana, Dominican Republic: Associa...

    2021

  29. [37]

    Yin, S.; Zhu, P.; Wu, L.; Gao, C.; and Wang, Z. 2024. GAMC: An Unsupervised Method for Fake News Detection Using Graph Autoencoder with Masking. Proceedings of the AAAI Conference on Artificial Intelligence, 38(1): 347--355

    2024

  30. [38]

    Yin, X.; Han, J.; and Yu, P. S. 2008. Truth Discovery with Multiple Conflicting Information Providers on the Web. IEEE Transactions on Knowledge and Data Engineering, 20(6): 796--808

    2008

  31. [39]

    Yuan, C.; Ma, Q.; Zhou, W.; Han, J.; and Hu, S. 2019. Jointly Embedding the Local and Global Relations of Heterogeneous Graph for Rumor Detection . In 2019 IEEE International Conference on Data Mining (ICDM), 796--805. Los Alamitos, CA, USA: IEEE Computer Society

    2019

  32. [40]

    Zhang, S.; Ma, X.; Duh, K.; and Van Durme, B. 2019. AMR Parsing as Sequence-to-Graph Transduction. In ACL, 80--94. Florence, Italy: ACL

    2019

  33. [41]

    Zhang, Y.; Trinh, L.; Cao, D.; Cui, Z.; and Liu, Y. 2023. Detecting Out-of-Context Multimodal Misinformation with interpretable neural-symbolic model. arXiv:2304.07633

    2023 arXiv

  34. [42]

    Zhou, X.; Wu, J.; and Zafarani, R. 2020. SAFE: Similarity-Aware Multi-modal Fake News Detection. In Lauw, H. W.; Wong, R. C.-W.; Ntoulas, A.; Lim, E.-P.; Ng, S.-K.; and Pan, S. J., eds., Advances in Knowledge Discovery and Data Mining, 354--367. Cham: Springer International Pu...

    2020

Pith tools

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