Pith. sign in

REVIEW 4 major objections 6 minor 46 references

MVAN: Multi-View Attention Networks for Fake News Detection on Social Media

T0 review · 4 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 MVAN detects fake news at 92–94% accuracy from a tweet and its retweet network.

desk verdict A reasonable BiGRU+GAT composition, but the headline accuracy gain is not established because the GCAN baseline is copied and the significance table is internally inconsistent. read the letter →

arxiv 2506.01627 v1 pith:4OVIF674 submitted 2025-06-02 cs.CL

Shiwen Ni , Jiawen Li , Hung-Yu Kao This is my paper
classification cs.CL
keywords fakenewsdetectionmulti-viewattentiongraphnetworkspropagationstructuretextsemanticinterpretabilityearlyTwitter
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 aims to show that fake news on Twitter can be detected from two sparse signals alone: the short text of the source tweet and the set of users who retweeted it, with no comment threads. It proposes MVAN, a multi-view network that reads the text with a bi-directional GRU and word attention, reads the retweet structure with a graph attention network over user profile features, and combines both views for classification. On the Twitter15 and Twitter16 datasets the model reports accuracies of 0.9234 and 0.9365, about 2.5% higher on average than the previous best model G-SEGA, with the differences reported as statistically significant at 0.90, 0.95, and 0.98 confidence levels. The paper also claims the attention mechanisms make decisions explainable by highlighting clue words in the tweet and suspicious users in the propagation chain, and that detection works early, holding near 91% accuracy when only part of the propagation has happened.

What carries the argument

The load-bearing mechanism is the pairing of two attention heads over two views. The text semantic attention network runs a bi-directional GRU over the tweet's word embeddings and computes a softmax weight per word, producing a weighted text vector. The propagation structure attention network runs a graph attention layer over the retweet graph, where each node is a retweeter described by 15 normalized Twitter profile features and attention is masked to first-order neighbors with multi-head averaging; it outputs a user-node representation. The prediction module concatenates the text and propagation vectors and classifies with softmax. The same attention weights double as explanations, since the model highlights the words and the users that most influenced the decision.

What would settle it

Retrain MVAN on the same Twitter15 and Twitter16 splits with the 15 profile features replaced by a random subset of the 38 crawled features, or by account age and follower count alone; if the accuracy gap over G-SEGA falls inside the reported standard deviation, the claimed driver of the gain is falsified. As a temporal check, train on older tweets and test on newer ones to see whether the user-metadata signal survives changes in Twitter's user population and verification rules.

Watch

Extended reading notes

Core claim

MVAN's central claim is that a short source tweet plus retweet-user metadata carries enough signal to classify a news item as true or fake, and that attending over both views explicitly outperforms methods that use richer content such as user comments. The model encodes the source tweet with word2vec and a two-layer BiGRU, applies text semantic attention to pool the hidden states, encodes the retweet graph node-wise with a two-layer multi-head graph attention network whose node features are 15 Twitter account attributes, and concatenates the two representations into a softmax classifier. Evaluated on Twitter15 and Twitter16, it reports 0.9234 and 0.9365 accuracy, improving on G-SEGA by roughly 3.06 and 2.03 percentage points, and ablations show the propagation view contributes the larger share of the gain. The attention weights are offered as explanations: words such as 'confirmed' mark true news, question marks mark fake news, and the most highly weighted early retweeters tend to be authoritative accounts for true news and newer, sparse-profile accounts for fake news.

Load-bearing premise

The load-bearing premise is that the 15 hand-selected Twitter profile features, with missing values filled by the mean of other users in the same propagation tree, capture enough of each user's credibility for the propagation structure to be informative; the paper does not test how sensitive the result is to this feature choice or imputation.

Editorial extensions

If this is right

  • Using only source text and retweet structure, the model reaches 0.9234 accuracy on Twitter15 and 0.9365 on Twitter16, beating the previous best reported system by about 2.5% on average.
  • Fake news can be detected early: with an early detection deadline, MVAN holds accuracy near 91%, so moderation could act before a story fully propagates.
  • The two attention views are complementary: ablations show that dropping either attention mechanism costs about 1% accuracy, while using only the propagation view costs about 9% and using only the text view costs about 3–4%.
  • The model gives per-case explanations: attention weights mark clue words in the tweet and suspicious users in the retweet chain, letting a human see why a particular item was flagged.
  • Since the model does not need user comments, it applies to realistic settings where only the source tweet and retweet user IDs are available.

Reading between the lines

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

  • A testable extension the paper leaves implicit: the same two-view attention recipe could be applied to other platform signals, such as retweet timing and follower overlap, and to other short-text classification tasks in the same comment-free setting, such as spam or coordinated-account detection.
  • If the result holds, the large contribution of the propagation view implies that social graph structure, not text, is the stronger veracity signal at this scale, so detectors should invest in user metadata and graph features rather than richer language models.
  • The reported attention patterns imply a simple content heuristic that could be checked directly: on these datasets, tweets containing words like 'confirmed' should be predominantly true and tweets containing explicit question marks predominantly fake, which a unigram baseline could verify.
  • The early-detection claim suggests that the first retweeters' profile features already carry most of the signal; a natural stress test is to hide the first few retweeters from the model and measure how much accuracy drops.
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 / 6 minor

Summary. The paper proposes MVAN, a multi-view attention network for fake news detection that uses only the source tweet text and the retweet propagation structure, without user comments. The model combines a BiGRU with a text semantic attention mechanism and a graph attention network over user features, and the two attention mechanisms are also used to provide word-level and user-level explanations. Experiments on Twitter15 and Twitter16 report accuracy improvements of roughly 2.5% on average over state-of-the-art baselines, together with ablation studies, early-detection experiments, and interpretability analyses.

Significance. If the empirical claims were rigorously supported, MVAN would be a practically useful detector for the realistic setting where only short source tweets and retweet structures are available, and the interpretability analysis would be a valuable addition. The paper also provides a plausible architectural combination of text and propagation attention. However, the current experimental validation has serious uncontrolled-comparison and statistical-reporting problems that prevent the central accuracy claim from being accepted as stated; the work has potential but needs a major revision with corrected experiments and statistics.

major comments (4)
  1. [Table 4 and Experimental Setup] The footnote to Table 4 states that GCAN's results are 'directly taken from the results shown in the original paper,' while the Experimental Setup says 'The results of the experiment are an average of ten times' and Table 5's caption asserts 'Each model we ran 10 times.' These statements are contradictory for GCAN, and no evidence is provided that the other baselines were rerun under MVAN's exact 70/30 split, preprocessing, and ten-seed protocol. Because the headline improvement over G-SEGA is only 2-3% in accuracy, a margin that can easily arise from different data splits or random seeds, the central empirical claim is not supported by the evidence as presented.
  2. [Table 5] Table 5 reports the same mean accuracy at three confidence levels with different ± values (e.g., SVM-BOW 0.6694 ± 0.052, ± 0.062, ± 0.073). A sample standard deviation does not change with confidence level, so the printed numbers are either mislabeled standard deviations or confusion between standard deviation and confidence-interval width. No t-statistics or p-values are reported, making the assertion that 'our model significantly outperforms all the baselines based on t-tests' unverifiable as printed.
  3. [Section 4.1 (Datasets) and Table 3] The user features were crawled after the fact via the Twitter API, not at the time of each tweet's propagation, and 15 of 38 features are selected manually with no sensitivity analysis. Missing user features for 5.4-7.2% of users are imputed with the mean of other users in the same propagation tree. The paper does not assess how feature temporality or imputation affects the results; if the crawled features reflect post-hoc account status rather than contemporaneous credibility, the reported margin over baselines could shrink or vanish.
  4. [Figure 3 (Ablation Study)] The ablation bar chart reports only mean accuracy without error bars or significance tests. The claims that removing the text or propagation attention costs about 1% and that removing the propagation structure entirely costs about 9% are not statistically supported; standard deviations or confidence intervals should be added, especially since the Table 5 discussion emphasizes statistical testing.
minor comments (6)
  1. [Introduction] The contribution list numbers two items as '(1)'; the second should be '(2)'.
  2. [Throughout] The model name appears inconsistently as 'MVAN', 'MV AN', and 'MA VN'; please standardize.
  3. [Experimental Setup and Table 4] The baseline is listed as 'G-SAGE' in the bullet list and 'G-SEGA' in Table 4; the correct name should be verified and used consistently.
  4. [Equation (21)] Equation (21) has a typo: it should be −(1−y)log(1−ŷ_f) rather than −(1−y)log(1−ŷ_f) with a subscript mismatch; the notation for the two predicted labels should be clarified.
  5. [Parameter Setting] The maximum text length L is defined in the model description but its value is never specified in the parameter settings.
  6. [Experimental Setup] The statement 'We followed GCAN [25] to split the datasets' is vague; the exact splitting procedure, including any random seed, should be reported for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: MVAN's accuracy claim rests on held-out labeled data; baseline-copy and significance-table issues are reproducibility problems, not definitional circularity.

full rationale

The paper's central claim is empirical: MVAN combines BiGRU text encoding with word-level attention and a GAT-based propagation structure encoder, is trained on labeled Twitter15/Twitter16 source tweets and retweet structures, and is evaluated on test splits. This is not circular by construction: the ground-truth labels are external to the model, the test set is separate, and no model parameter is fitted to the test labels. The attention-based 'explanations' are post hoc visualizations of learned weights rather than predictions derived from a first-principles theory, so they cannot render the claim circular either. The authors cite their own prior work (e.g., references [20], [26], [27], [41]) only as related background and technical building blocks; none of these citations supplies a load-bearing premise from which the reported accuracy gain follows, and the model is not defined in terms of the outcome it predicts. The main validity problems are not circularity. Table 4 states that GCAN's results are 'directly taken from the results shown in the original paper,' while Table 5's caption claims 'Each model we ran 10 times'; these statements conflict for GCAN, and Table 5 reports different +/- values for the same mean at different confidence levels, which is internally inconsistent as printed. Those are reproducibility, fairness, and reporting issues, not definitional circularity, and they do not make the trained model's predictions equivalent to its inputs. I therefore find no significant circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The ledger lists the hand-set hyperparameters and domain assumptions the central accuracy claim depends on. None of these are fitted to the test set, but they are not varied in a sensitivity analysis, so the reader cannot tell how robust the reported margins are. No new entities are postulated.

free parameters (6)
  • BiGRU hidden size = 300
    Set to 300 in Section 4.2; no sensitivity analysis reported.
  • Number of BiGRU layers = 2
    Set to 2 in Section 4.2; no sensitivity analysis reported.
  • Number of GAT layers = 2
    Set to 2 in Section 4.2; no sensitivity analysis reported.
  • Number of attention heads = 5
    Set to 5 in Section 4.2; follows common GAT practice.
  • Selected user features = 15 of 38 crawled
    Hand-picked from prior work [1,8,44]; feature selection could affect results.
  • Maximum text length L = not stated
    Zero-padding is used but the maximum length is not reported; text representation depends on this choice.
assumptions (5)
  • domain assumption Pre-trained GoogleNews word2vec embeddings capture enough semantic information from very short tweets (average 13 words).
    Used in the text encoder; no fine-tuning is described.
  • domain assumption First-order graph attention over retweet neighbors captures the veracity-relevant structure.
    The propagation attention network computes attention over first-order neighbors only (Section 3.3).
  • domain assumption The random 70/30 split with 10 runs is a fair evaluation protocol.
    Comparison baselines may have used different splits; GCAN results are copied from the original paper.
  • domain assumption Mean imputation for missing user features does not bias results.
    Missing users (5.4 to 7.2 percent) are filled with the mean of other users in the same propagation structure (Section 4.1).
  • domain assumption Reducing the four-class rumor datasets to binary true/fake preserves the detection task.
    The paper drops non-rumor and unverified classes without discussing the impact on difficulty or comparability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MVAN: Multi-View Attention Networks for Fake News Detection on Social Media." pith.science (2026). https://pith.science/paper/4OVIF674

@misc{pith2026250601627,
  author       = {Pith},
  title        = {Pith review of: MVAN: Multi-View Attention Networks for Fake News Detection on Social Media},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4OVIF674}},
  note         = {Machine review of arXiv:2506.01627}
}
read the original abstract

Fake news on social media is a widespread and serious problem in today's society. Existing fake news detection methods focus on finding clues from Long text content, such as original news articles and user comments. This paper solves the problem of fake news detection in more realistic scenarios. Only source shot-text tweet and its retweet users are provided without user comments. We develop a novel neural network based model, \textbf{M}ulti-\textbf{V}iew \textbf{A}ttention \textbf{N}etworks (MVAN) to detect fake news and provide explanations on social media. The MVAN model includes text semantic attention and propagation structure attention, which ensures that our model can capture information and clues both of source tweet content and propagation structure. In addition, the two attention mechanisms in the model can find key clue words in fake news texts and suspicious users in the propagation structure. We conduct experiments on two real-world datasets, and the results demonstrate that MVAN can significantly outperform state-of-the-art methods by 2.5\% in accuracy on average, and produce a reasonable explanation.

Figures

Figures reproduced from arXiv: 2506.01627 by the authors.

Figure 1
Figure 1. Propagation structure of Twitter on social media. A tweet can be reposted by multiple people, and a person can also [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Architecture of the MVAN model. The two attention mechanisms represent two different kinds of information. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. MVAN ablation analysis in accuracy. The results [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Early fake news detection results. show that our model has a significant performance improve￾ment on the two public datasets. MVAN can better represent text and propagation structural information, thereby improv￾ing the accuracy of fake news detection. Ablation Study T…
Figure 5
Figure 5. Figure 5: Distribution of statistics of word attention weight value. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Evidence words through word cloud visualiza [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Real case analysis of true news and fake news. The key clue words in the source tweet are highlighted by text [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 41 canonical work pages

  1. [1]

    Informa- tion credibility on twitter.In Proceedings of the 20th international conference on World wide web.2011, pp

    Castillo, C., Mendoza, M., and Poblete, B. Informa- tion credibility on twitter.In Proceedings of the 20th international conference on World wide web.2011, pp. 675-684

    work page 2011

  2. [2]

    Rumor has it: Identifying misinformation in mi- croblogs.In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing

    Qazvinian, V ., Rosengren, E., Radev, D., and Mei, Q. Rumor has it: Identifying misinformation in mi- croblogs.In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing. 2011, pp. 1589-1599

    work page 2011

  3. [3]

    Assessing the credibility of claims on the web.In Proceedings of the 26th International Confer- ence on World Wide Web Companion.2017, pp

    Popat, K. Assessing the credibility of claims on the web.In Proceedings of the 26th International Confer- ence on World Wide Web Companion.2017, pp. 735- 739

    work page 2017

  4. [4]

    Automatic de- tection of rumor on sina weibo.In Proceedings of the ACM SIGKDD workshop on mining data semantics

    Yang, F., Liu, Y ., Yu, X., and Yang, M. Automatic de- tection of rumor on sina weibo.In Proceedings of the ACM SIGKDD workshop on mining data semantics. 2012, pp. 1-7

    work page 2012

  5. [5]

    Epidemiological modeling of news and ru- mors on twitter.In Proceedings of the 7th workshop on social network mining and analysis.2013, pp

    Jin, F., Dougherty, E., Saraf, P., Cao, Y ., and Ramakr- ishnan, N. Epidemiological modeling of news and ru- mors on twitter.In Proceedings of the 7th workshop on social network mining and analysis.2013, pp. 1-9

    work page 2013

  6. [6]

    Lever- aging the implicit structure within social media for emergent rumor detection,

    J. Sampson, F. Morstatter, L. Wu, and H. Liu, “Lever- aging the implicit structure within social media for emergent rumor detection,”in Proceedings of the 25th ACM International on Conference on Information and Knowledge Management.ACM, 2016, pp. 2377–2382

    work page 2016

  7. [7]

    Detect rumors in microblog posts using propagation structure via kernel learning,

    J. Ma, W. Gao, and K.-F. Wong, “Detect rumors in microblog posts using propagation structure via kernel learning,”in Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Vol- ume 1: Long Papers), 2017, pp. 708–717

    work page 2017

  8. [8]

    Early detection of fake news on social media through propagation path classifica- tion with recurrent and convolutional networks,

    Y . Liu and Y .-F. B. Wu, “Early detection of fake news on social media through propagation path classifica- tion with recurrent and convolutional networks,”in Thirty-Second AAAI Conference on Artificial Intelli- gence, 2018

    work page 2018

Show all 46 references
  1. [9]

    CED: credible early detection of social media ru- mors.IEEE Transactions on Knowledge and Data En- gineering.2019

    Song, C., Yang, C., Chen, H., Tu, C., Liu, Z., and Sun, M. CED: credible early detection of social media ru- mors.IEEE Transactions on Knowledge and Data En- gineering.2019

    2019

  2. [10]

    J., Wong, K

    Ma, J., Gao, W., Mitra, P., Kwon, S., Jansen, B. J., Wong, K. F., and Cha, M. Detecting rumors from mi- croblogs with recurrent neural networks.Proceedings of the 25th International Joint Conference on Artificial Intelligence.2016

    2016

  3. [11]

    Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., and Philip, S. Y . A comprehensive survey on graph neural networks.IEEE Transactions on Neural Networks and Learning Systems.2020

    2020

  4. [12]

    Wu, K., Yang, S., and Zhu, K. Q. 2015. False rumors detection on Sina Weibo by propagation structures. 2015 IEEE 31st International Conference on Data En- gineering,2015-May, 651–662

    2015

  5. [13]

    Rath, B., Gao, W., Ma, J., and Srivastava, J. 2017. From retweet to believability: Utilizing trust to identify rumor spreaders on twitter.Proceedings of the 2017 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining,ASONAM 2017, 179–186

    2017

  6. [14]

    A con- volutional approach for misinformation identification

    Yu, F., Liu, Q., Wu, S., Wang, L., and Tan, T. A con- volutional approach for misinformation identification. Proceedings of the 26th International Joint Conference on Artificial Intelligence.2017

    2017

  7. [15]

    Call atten- tion to rumors: Deep attention based recurrent neu- ral networks for early rumor detection.In Pacific-Asia conference on knowledge discovery and data mining

    Chen, T., Li, X., Yin, H., and Zhang, J. Call atten- tion to rumors: Deep attention based recurrent neu- ral networks for early rumor detection.In Pacific-Asia conference on knowledge discovery and data mining. 2018, pp. 40-52

    2018

  8. [16]

    Liu, Y ., and Wu, Y . F. B. Early detection of fake news on social media through propagation path clas- sification with recurrent and convolutional networks. In Thirty-second AAAI conference on artificial intelli- gence.2018

    2018

  9. [17]

    Fake news identification on twitter with hybrid cnn and rnn mod- els.In Proceedings of the 9th international conference on social media and society.2018, pp

    Ajao, O., Bhowmik, D., and Zargari, S. Fake news identification on twitter with hybrid cnn and rnn mod- els.In Proceedings of the 9th international conference on social media and society.2018, pp. 226-230

    2018

  10. [18]

    Yu, F., Liu, Q., Wu, S., Wang, L., and Tan, T Attention-based convolutional approach for misinfor- mation identification from massive and noisy mi- croblog posts.computers & security.2019, 83, 106-121

    2019

  11. [19]

    defend: Explainable fake news detec- tion.In:Proceedings of the 25th ACM SIGKDD Inter- national Conference on Knowledge Discovery & Data Mining.2019

    SHU, Kai, et al. defend: Explainable fake news detec- tion.In:Proceedings of the 25th ACM SIGKDD Inter- national Conference on Knowledge Discovery & Data Mining.2019. pp. 395-405

    2019

  12. [20]

    Sujana, Yudianto, Jiawen Li, and Hung-Yu Kao. Ru- mor Detection on Twitter Using Multiloss Hierarchi- cal BiLSTM with an Attenuation Factor.Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International...

    2020

  13. [21]

    Ma, J., Gao, W., and Wong, K. F. Rumor detection on twitter with tree-structured recursive neural networks. Association for Computational Linguistics.2018

    2018

  14. [22]

    Fake news detection on so- cial media using geometric deep learning,

    F.Monti,F.Frasca,D.Eynard,D.Mannion,and M.M.Bronstein,“Fake news detection on so- cial media using geometric deep learning,” 2019, arXiv:1902.06673. [Online]. Available: http://arxiv.org/abs/1902.06673

    2019 arXiv

  15. [23]

    Graph-based Rumour detection for so- cial media,

    T. T. Nguyen, “Graph-based Rumour detection for so- cial media,” Infosci., Tech. Rep., 2019. [Online]. Avail- able: https://infoscience.epfl.ch/record/265745

    2019

  16. [24]

    ”Rumor Detection on Social Me- dia with Bi-Directional Graph Convolutional Net- works.”Proceedings of the AAAI Conference on Arti- ficial Intelligence.V ol

    Bian, Tian, et al. ”Rumor Detection on Social Me- dia with Bi-Directional Graph Convolutional Net- works.”Proceedings of the AAAI Conference on Arti- ficial Intelligence.V ol. 34. No. 01. 2020

    2020

  17. [25]

    Lu, Yi-Ju, and Cheng-Te Li. ”GCAN: Graph-aware Co-Attention Networks for Explainable Fake News Detection on Social Media.”Proceedings of The 58th Annual Meeting of the Association for Computational Linguistics.ACL. 2020

    2020

  18. [26]

    Ex- ploiting Microblog Conversation Structures to Detect Rumors.Proceedings of the 28th International Con- ference on Computational Linguistics.2020

    Li, Jiawen, Yudianto Sujana, and Hung-Yu Kao. Ex- ploiting Microblog Conversation Structures to Detect Rumors.Proceedings of the 28th International Con- ference on Computational Linguistics.2020

    2020

  19. [27]

    ”Birds of a Feather Rumor Together? Exploring Homogeneity and Conversation Structure in Social Media for Rumor De- tection.”IEEE Access.2020

    Li, Jiawen, Shiwen Ni, and Hung-Yu Kao. ”Birds of a Feather Rumor Together? Exploring Homogeneity and Conversation Structure in Social Media for Rumor De- tection.”IEEE Access.2020

    2020

  20. [28]

    R., and Jain, L

    Medsker, L. R., and Jain, L. C. . Recurrent neural net- works.Design and Applications,2001, 5

    2001

  21. [29]

    ”Learning long-term dependencies with gradient de- scent is difficult.”IEEE transactions on neural net- works5.2 (1994): 157-166

    Bengio, Yoshua, Patrice Simard, and Paolo Frasconi. ”Learning long-term dependencies with gradient de- scent is difficult.”IEEE transactions on neural net- works5.2 (1994): 157-166

    1994

  22. [30]

    LSTM can solve hard long time lag problems.In: Advances in neural information processing systems

    HOCHREITER, Sepp; SCHMIDHUBER, J ¨urgen. LSTM can solve hard long time lag problems.In: Advances in neural information processing systems

  23. [31]

    ”Speech recognition with deep recurrent neu- ral networks.”2013 IEEE international conference on acoustics, speech and signal processing.IEEE, 2013

    Graves, Alex, Abdel-rahman Mohamed, and Geoffrey Hinton. ”Speech recognition with deep recurrent neu- ral networks.”2013 IEEE international conference on acoustics, speech and signal processing.IEEE, 2013

    2013

  24. [32]

    ”Empirical evaluation of gated recurrent neural networks on sequence model- ing.”arXiv preprint arXiv:1412.3555(2014)

    Chung, Junyoung, et al. ”Empirical evaluation of gated recurrent neural networks on sequence model- ing.”arXiv preprint arXiv:1412.3555(2014)

    2014 arXiv

  25. [33]

    ”A neural probabilistic language model.”Journal of machine learning research3.Feb (2003): 1137-1155

    Bengio, Yoshua, et al. ”A neural probabilistic language model.”Journal of machine learning research3.Feb (2003): 1137-1155

    2003

  26. [34]

    ”Recurrent models of visual attention.”Advances in neural information processing systems27 (2014): 2204-2212

    Mnih, V olodymyr, Nicolas Heess, and Alex Graves. ”Recurrent models of visual attention.”Advances in neural information processing systems27 (2014): 2204-2212

    2014

  27. [35]

    ”Neural machine translation by jointly learning to align and translate.”arXiv preprint arXiv:1409.0473(2014)

    Bahdanau, Dzmitry, Kyunghyun Cho, and Yoshua Bengio. ”Neural machine translation by jointly learning to align and translate.”arXiv preprint arXiv:1409.0473(2014)

    2014 arXiv

  28. [36]

    ”Attention is all you need.”Advances in neural information processing sys- tems.2017

    Vaswani, Ashish, et al. ”Attention is all you need.”Advances in neural information processing sys- tems.2017

    2017

  29. [37]

    ”Graph neural networks: A re- view of methods and applications.”arXiv preprint arXiv:1812.08434 (2018)

    Zhou, Jie, et al. ”Graph neural networks: A re- view of methods and applications.”arXiv preprint arXiv:1812.08434 (2018)

    2018 arXiv

  30. [38]

    Graph attention networks,

    P. Velickovic, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y . Bengio, “Graph attention networks,”ICLR 2018,2018

    2018

  31. [39]

    ”Rumor detection on twitter with tree-structured recursive neu- ral networks.”Association for Computational Linguis- tics,2018

    42.Ma, Jing, Wei Gao, and Kam-Fai Wong. ”Rumor detection on twitter with tree-structured recursive neu- ral networks.”Association for Computational Linguis- tics,2018

    2018

  32. [40]

    ”Stance detection with bidirectional conditional encoding.” 2016

    43.Augenstein, Isabelle, et al. ”Stance detection with bidirectional conditional encoding.” 2016. [Online]. Available: arXiv preprint arXiv:1606.05464

    2016 arXiv

  33. [41]

    44.Chen, Yi-Chin, Zhao-Yang Liu, and Hung-Yu Kao. ”IKM at SemEval-2017 Task 8: Convolutional neural networks for stance detection and rumor verification.” Proceedings of the 11th International Workshop on Se- mantic Evaluation (SemEval-2017).2017

    2017

  34. [42]

    ”Csi: A hybrid deep model for fake news detection.” Proceedings of the 2017 ACM on Conference on Infor- mation and Knowledge Management.2017

    45.Ruchansky, Natali, Sungyong Seo, and Yan Liu. ”Csi: A hybrid deep model for fake news detection.” Proceedings of the 2017 ACM on Conference on Infor- mation and Knowledge Management.2017

    2017

  35. [43]

    Liu, Yang, and Yi-Fang Brook Wu. ”Early detection of fake news on social media through propagation path classification with recurrent and convolutional networks.”Thirty-second AAAI conference on artificial intelligence.2018

    2018

  36. [44]

    Rumor detection with hierarchical so- cial attention network.Proceedings of the 27th ACM International Conference on Information and Knowl- edge Management.2018

    GUO, Han, et al. Rumor detection with hierarchical so- cial attention network.Proceedings of the 27th ACM International Conference on Information and Knowl- edge Management.2018. p. 943-951

    2018

  37. [45]

    ”Spatial temporal graph convolutional networks for skeleton- based action recognition.”Proceedings of the AAAI conference on artificial intelligence.V ol

    Yan, Sijie, Yuanjun Xiong, and Dahua Lin. ”Spatial temporal graph convolutional networks for skeleton- based action recognition.”Proceedings of the AAAI conference on artificial intelligence.V ol. 32. No. 1. 2018

    2018

  38. [46]

    Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting.In- ternational Conference on Learning Representations

    LI, Yaguang, et al. Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting.In- ternational Conference on Learning Representations. 2018

    2018

Pith tools

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