Pith. sign in

REVIEW 4 major objections 6 minor 42 references

MisinfoTeleGraph: Network-driven Misinformation Detection for German Telegram Messages

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Incorporating Telegram's forwarding network lifts misinformation detection from 0.69 to 0.95 MCC on a new German-language dataset.

desk verdict Useful new German Telegram graph dataset, but the headline GNN-vs-text comparison is unproven due to duplicate leakage and missing evaluation rigor. read the letter →

arxiv 2506.22529 v1 pith:YQIQZ3SF submitted 2025-06-27 cs.CL

Lu Kalkbrenner , Veronika Solopova , Steffen Zeiler , Robert Nickel , Dorothea Kolossa This is my paper
classification cs.CL
keywords misinformationdetectionTelegramgraphneuralnetworksSAGEGermanlanguageweaksupervisionmessageforwardingM3embeddings
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 misinformation on Telegram can be detected more reliably when the model sees not just message text but the network structure of who forwards what to whom. To test this, it builds a new German-language dataset of over 5 million public Telegram messages with weak labels from fact-check and news similarity and a smaller set of manually verified labels. A GraphSAGE graph neural network using forwarding edges reaches a Matthews correlation coefficient of 0.95 on the test set, compared with 0.69 for a text-only baseline. The paper also reports that adding view and subscriber counts helps slightly, and that training on automatically generated weak labels performs about as well as training on manually annotated labels. If the result generalizes, network-driven detection is a practical route for monitoring low-moderation platforms like Telegram.

What carries the argument

The load-bearing mechanism is a heterogeneous graph with two node types, Telegram channels and messages, and two edge types: IS_PART_OF (message posted in a channel) and FORWARDED (message forwarded across channels). Messages are duplicated when forwarded so the original channel is preserved. Node features are M3 text embeddings of message or channel text concatenated with views or subscriber counts. GraphSAGE with LSTM aggregation propagates these features over the graph, and four layers capture the four-hop structure needed to reach all messages in a channel and the channels that forward them. The central comparison is this graph model against a text-only model that classifies the same M3 embeddings without any network information.

What would settle it

Cluster messages by semantic similarity or by forwarding chains before splitting the dataset, so near-duplicate and forwarded copies never cross the train/test boundary; if the graph model's MCC gap over the text baseline collapses or shrinks materially, the reported advantage came from label leakage rather than network signal.

Watch

Extended reading notes

Core claim

The paper's central claim is that incorporating message-forwarding relationships into a graph neural network improves misinformation detection over text-only classification on Telegram. On the MisinfoTeleGraph dataset, the graph-based GraphSAGE model with LSTM aggregation and four layers reaches an MCC of 0.95, precision and recall near or above 0.9 for both classes, while the text-only baseline reaches MCC 0.691. The authors attribute the advantage to the four-hop neighborhood that lets a message aggregate information from its own channel and all channels it is forwarded to. They also find that view and subscriber counts add a small performance boost, and that weak supervision via semantic similarity to fact-checks and news articles does not hurt performance compared to manual labels.

Load-bearing premise

The evaluation assumes the train/test split keeps forwarded copies and near-identical messages out of the test set; the paper's limitations section says similar messages may appear in both training and test sets.

Editorial extensions

If this is right

  • Network structure helps: the GraphSAGE forwarding-aware model beats the text-only model on every reported metric, with MCC 0.950 versus 0.691.
  • Weak labels are usable: training on automatically generated semantic-similarity labels performs nearly as well as training on manually verified labels (MCC 0.950 versus 0.923), with calibration error below 0.1.
  • Metadata matters: adding view and subscriber counts to node features improves MCC from 0.901 to 0.950.
  • The dataset supports network analysis: centrality measures such as forward-degree centrality separate content-creating channels from content-amplifying channels.
  • Four GraphSAGE layers fit this graph's structure because a message's four-hop neighborhood covers its own channel and the channels it is forwarded to.

Reading between the lines

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

  • If the reported advantage is real, the same forwarding-aware architecture could transfer to other low-moderation platforms where propagation structure is observable, but a practical deployment would need a continuously updated graph of current forwarding events.
  • The limitation the authors flag about similar messages crossing splits suggests a concrete follow-up: re-running the same comparison under cluster-based splitting would show how much of the 0.95 versus 0.69 gap is generalization rather than memorization.
  • The observed cross-lingual matching of English and Russian messages to German claims hints that M3 embeddings can be used to build multilingual misinformation graphs without per-language annotation, though the logical-entailment errors the authors describe would limit that use.
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 introduces MisinfoTeleGraph, a German-language Telegram graph dataset containing over 5 million messages from 13,845 channels, with weak labels derived from M3-embedding similarity to fact-checks/news articles and strong labels from manual annotation. The authors construct a heterogeneous graph with message-channel and message-forwarding edges, and compare a GraphSAGE model with LSTM aggregation against a text-only M3-embedding classifier. They report that the graph-based model substantially outperforms the text baseline (MCC 0.950 vs. 0.691), and additionally analyze the effect of view/subscriber counts, weak versus strong labels, and network centrality measures. The claimed central result is that incorporating Telegram forwarding structure improves misinformation detection over a text-only baseline.

Significance. If the central claim holds, this is a useful contribution: it provides a new German-language Telegram misinformation dataset, an open code base, reproducible baselines, and an application of graph neural networks to a platform and language that are underrepresented in misinformation research. The dataset scale and the inclusion of both weak and strong labels are valuable assets. However, the evaluation protocol has load-bearing weaknesses: the train/test split is not documented, forwarded duplicates can span the split, hyperparameters are selected on the test set, and the reported dataset statistics are internally inconsistent. These issues directly affect whether the reported 25.9-point MCC gap reflects generalization or leakage, so the central claim is not yet established to the standard required by the paper's conclusions.

major comments (4)
  1. [§3.3, §5.4, Table 4; Limitations] The central comparison in Table 4 is vulnerable to duplication leakage. Section 3.3 states that messages are duplicated when forwarded and linked by FORWARDED edges, so identical or near-identical text appears in multiple nodes. The paper reports no train/test split method, no random seed, and no cluster-level or connected-component-level separation. With a random node-level split, a forwarded copy can land in the test set while the original and other copies remain in the training graph; a four-layer GraphSAGE model can then aggregate features from those training copies through forwarding and channel edges, whereas the text-only baseline cannot exploit this structure. The Limitations section concedes that 'similar messages may appear in both the training and test sets.' Therefore the reported 25.9-point MCC gap may reflect memorization of near-duplicates rather than generalization to unseen misinformation.
  2. [§5.3, Table 3, Figure 5] The final architecture is selected using test-set metrics. Table 3 and Figure 5 report test precision, recall, F1, and MCC for different aggregators, numbers of GraphSAGE layers, and numbers of epochs, and the authors then adopt the best-performing configuration for the main experiments in Tables 4-6. No validation split is described. This test-set-based model selection can inflate the reported MCC of 0.950, and the absence of repeated runs, error bars, or significance tests makes it impossible to assess whether the gap over the text baseline is stable.
  3. [Table 1, §3.2, Limitations] The dataset statistics are internally inconsistent and prevent reconstruction of the evaluation data. Table 1 reports 742 weakly linked message-claim pairs but lists 110 factual, 632 misinformation, and 542 'other' weak pairs, whose sum is 1,284. Section 3.2 reports 868 weakly annotated and 589 strongly annotated pairs, while the Limitations section reports 873 weak pairs and 651 strong pairs. Table 1 reports 651 strong pairs. These contradictions must be resolved, and the precise train/test sizes and label distributions must be stated, before the reported results can be reproduced or trusted.
  4. [§3.2, §3.3, §5.6] There is a partial circularity between the weak labels and the text features. Weak labels are generated by thresholding M3-embedding similarity between messages and fact-check claims (Section 3.2), and the node features used by both the GNN and the text baseline are the same M3 embeddings (Section 3.3). Training on weak labels can therefore partly reproduce the matching rule that produced the labels, rather than learning a general misinformation detector. The strong labels and manual test annotation provide some independence, but the main GNN-versus-text comparison in Table 4 is reported on the weak-label setup; the paper should also report the equivalent comparison trained and evaluated on the strong/manually annotated labels to separate network-structure gains from weak-label circularity.
minor comments (6)
  1. [Abstract and Title] The dataset name is written as both 'MisinfoTeleGraph' and 'Misinfo-TeleGraph' across the abstract, title, and body; please unify the spelling.
  2. [Figure 2] The caption contains a typo: 'Pizzzagate' should be 'Pizzagate'.
  3. [§3.2] The source is referred to as 'WoldNewsAPI' but the URL is worldnewsapi.com; the spelling should be corrected to 'WorldNewsAPI'.
  4. [Table 1] The role of the 'other' class in weak pairs is unclear: the downstream model is binary (Sigmoid output for factual vs. misinformation), so the paper should explain what the 542 'other' pairs are and why they are excluded from training and evaluation.
  5. [§3.4] The learning rate scheduler is described as starting at 1e-3 and ending at 1e-5 'after 100 iterations,' but the main experiments use 10 epochs; please clarify whether the scheduler decays per epoch or per iteration.
  6. [Abstract] The abstract writes '13.845' with a decimal comma, while the body uses '13,845'; please use consistent thousands separators.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the graph-vs-text comparison is an empirical benchmark; weak-label/text-embedding overlap is a weak-supervision design choice, and the acknowledged duplication leakage is a data-splitting correctness risk, not a derivation that reduces to its inputs.

full rationale

The paper's central claim is empirical: GraphSAGE with LSTM aggregation outperforms a text-only M3 classifier on MCC and F1. The weak labels are produced by M3-embedding similarity to external fact-check/newspaper claims, and the same M3 embeddings are used as text features; this overlap is a standard weak-supervision pipeline rather than a circular derivation, because the labels are anchored to an external knowledge base and to manual annotation, and the graph signal (FORWARDED and IS_PART_OF edges) is not determined by the embeddings. The Limitations section explicitly concedes that similar messages may appear in both training and test sets because of forwarding and minor text modifications; this is a data-leakage and generalization-risk issue, not a case where a prediction is equivalent to its inputs by construction. The internal inconsistencies in dataset counts (e.g., Table 1 class counts summing to 1284 weak pairs versus the stated 742 total, and 868 versus 873 weak pairs across sections) are reporting errors rather than circular reasoning. No load-bearing self-citation or imported uniqueness theorem is present; the one self-citation (Solopova et al., 2024) supports a general related-work statement and is not load-bearing. The comparison to external benchmarks such as Mahmud et al. (2022) and Nielsen and McConville (2022) is independent of the paper's fitted values, and the manual strong-label test set provides an external check. Therefore no step in the derivation chain reduces to its own inputs; the appropriate finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the weak-labeling pipeline, the specific GNN configuration, and external data sources. Most assumptions are domain choices about what the Telegram graph and fact-check similarity mean; none are formal mathematical axioms. No physical entities are invented.

free parameters (5)
  • similarity threshold = 0.7
    Selected by comparing precision of weak message-claim matching; controls which messages receive weak labels (Section 3.2).
  • GraphSAGE layers = 4
    Best test metrics; only up to 4 layers tested due to hardware (Section 5.3).
  • aggregator = LSTM
    Outperformed mean aggregation on test metrics (Table 3).
  • training epochs = 10
    Used for final experiments due to hardware limits (Section 5.3).
  • learning rate schedule and weight decay = 1e-3 to 1e-5, weight decay 1e-5
    Fixed hyperparameters, no sensitivity analysis (Section 3.4).
assumptions (4)
  • domain assumption Telegram forwarding edges are a meaningful propagation signal; duplicated messages with FORWARDED edges preserve channel origin.
    Graph construction and neighborhood definition rely on this (Section 3.3).
  • domain assumption M3-embedding similarity to fact-check claims is a valid weak label source at threshold 0.7.
    Weak labels are inherited from matched claims (Section 3.2).
  • domain assumption Manual annotation of the strong set is reliable ground truth.
    No inter-annotator agreement is reported (Sections 3.2, 5.6).
  • domain assumption Data4Transparency data accurately represents German public Telegram channels and messages.
    All graph and label statistics derive from this third-party collection (Section 3.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of MisinfoTeleGraph: Network-driven Misinformation Detection for German Telegram Messages." pith.science (2026). https://pith.science/paper/YQIQZ3SF

@misc{pith2026250622529,
  author       = {Pith},
  title        = {Pith review of: MisinfoTeleGraph: Network-driven Misinformation Detection for German Telegram Messages},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YQIQZ3SF}},
  note         = {Machine review of arXiv:2506.22529}
}
read the original abstract

Connectivity and message propagation are central, yet often underutilized, sources of information in misinformation detection -- especially on poorly moderated platforms such as Telegram, which has become a critical channel for misinformation dissemination, namely in the German electoral context. In this paper, we introduce Misinfo-TeleGraph, the first German-language Telegram-based graph dataset for misinformation detection. It includes over 5 million messages from public channels, enriched with metadata, channel relationships, and both weak and strong labels. These labels are derived via semantic similarity to fact-checks and news articles using M3-embeddings, as well as manual annotation. To establish reproducible baselines, we evaluate both text-only models and graph neural networks (GNNs) that incorporate message forwarding as a network structure. Our results show that GraphSAGE with LSTM aggregation significantly outperforms text-only baselines in terms of Matthews Correlation Coefficient (MCC) and F1-score. We further evaluate the impact of subscribers, view counts, and automatically versus human-created labels on performance, and highlight both the potential and challenges of weak supervision in this domain. This work provides a reproducible benchmark and open dataset for future research on misinformation detection in German-language Telegram networks and other low-moderation social platforms.

Figures

Figures reproduced from arXiv: 2506.22529 by the authors.

Figure 1
Figure 1. Weak Supervision using M3-embeddings and a knowledge base. A source text (on the left) is compared to [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Example of a Social Network Graph. The example shows the “Pizzzagate Archiv“ Telegram channel [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. GraphSAGE model architecture. The network [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Text-based model architecture. Based on the text data of the messages to be classified, M3- embeddings are computed (M3). The embeddings are then classified by a fully connected layer (FC) and a Sigmoid function (S). 5 Results 5.1 Metrics As standard evaluation metrics…
Figure 5
Figure 5. Figure 5: Test set metrics across two training configurations: (a) different numbers of training epochs using LSTM [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 25 canonical work pages

  1. [1]

    Google Fact Check Tools API. 2024. Google fact check tools api. https://developers.google.com/fact-check/tools/api/. Accessed: 2025-04-07

    work page 2024

  2. [2]

    Jason Baumgartner, Savvas Zannettou, Megan Squire, and Jeremy Blackburn. 2020. https://ojs.aaai.org/index.php/ICWSM/article/view/7348 The pushshift telegram dataset . In Proceedings of the Fourteenth International AAAI Conference on Web and Social Media, ICWSM 2020, Held Virtually, Original Venue: Atlanta, Georgia, USA, June 8-11, 2020 , pages 840--847. A...

    work page 2020

  3. [3]

    Alexandre Bovet and Peter Grindrod. 2020. https://doi.org/10.13140/RG.2.2.16700.05764 The activity of the far right on telegram . ResearchGate preprint, pages 1--19

    work page arXiv 2020

  4. [4]

    Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. 2024. https://doi.org/10.48550/ARXIV.2402.03216 BGE m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation . CoRR, abs/2402.03216

    arXiv doi:10.48550/arxiv.2402.03216 2024

  5. [5]

    Kousik Das, Sovan Samanta, and Madhumangal Pal. 2018. https://doi.org/10.1007/S13278-018-0493-2 Study on centrality measures in social networks: a survey . Soc. Netw. Anal. Min., 8(1):13

    work page doi:10.1007/s13278-018-0493-2 2018

  6. [6]

    Data4Transparency. 2024. Data4transparency. https://data4transparency.com/. Accessed: 2024-08-17

    work page 2024

  7. [7]

    Bruce Croft

    Mostafa Dehghani, Hamed Zamani, Aliaksei Severyn, Jaap Kamps, and W. Bruce Croft. 2017. https://doi.org/10.1145/3077136.3080832 Neural ranking models with weak supervision . In Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, Shinjuku, Tokyo, Japan, August 7-11, 2017 , pages 65--74. ACM

    arXiv 2017

  8. [8]

    Yu, and Lichao Sun

    Yingtong Dou, Kai Shu, Congying Xia, Philip S. Yu, and Lichao Sun. 2021. https://doi.org/10.1145/3404835.3462990 User preference-aware fake news detection . In SIGIR '21: The 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, Virtual Event, Canada, July 11-15, 2021 , pages 2051--2055. ACM

    arXiv 2021

Show all 42 references
  1. [9]

    Goodfellow, Jonathon Shlens, and Christian Szegedy

    Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. 2015. https://arxiv.org/abs/1412.6572 Explaining and harnessing adversarial examples . In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings

    2015 arXiv

  2. [10]

    Tarek Hamdi, Hamda Slimi, Ibrahim Bounhas, and Yahya Slimani. 2020. https://doi.org/10.1007/978-3-030-36987-3\_17 A hybrid approach for fake news detection in twitter based on user features and graph embedding . In Distributed Computing and Internet Technology - 16th Internati...

    2020 doi

  3. [11]

    Hamilton, Zhitao Ying, and Jure Leskovec

    William L. Hamilton, Zhitao Ying, and Jure Leskovec. 2017. https://proceedings.neurips.cc/paper/2017/hash/ \\ 5dd9db5e033da9c6fb5ba83c7a7ebea9-Abstract.html Inductive representation learning on large graphs . In Advances in Neural Information Processing Systems 30: Annual Conf...

    2017

  4. [12]

    Yi Han, Shanika Karunasekera, and Christopher Leckie. 2020. https://arxiv.org/abs/2007.03316 Graph neural networks with continual learning for fake news detection from social media . CoRR, abs/2007.03316

    2020 arXiv

  5. [13]

    Srinivas, Gagan K

    Srinidhi Hiriyannaiah, A.M.D. Srinivas, Gagan K. Shetty, Siddesh G.M., and K.G. Srinivasa. 2020. https://doi.org/10.1016/B978-0-12-818699-2.00004-4 Chapter 4 - a computationally intelligent agent for detecting fake news using generative adversarial networks . In Siddhartha Bha...

    2020 doi

  6. [14]

    Josef Holnburger. 2023. https://cemas.io/publikationen/telegram-chronologie-einer-radikalisierung/2023-03-22_CeMAS_Telegram_Chronologie_einer_Radikalisierung.pdf Chronologie einer Radikalisierung - Wie Telegram zur wichtigsten Plattform für Verschwörungsideologien und Rechtsex...

    2023

  7. [15]

    Howard, Bharath Ganesh, Dimitra Liotsiou, John Kelly, and Camille François

    Philip N. Howard, Bharath Ganesh, Dimitra Liotsiou, John Kelly, and Camille François. 2019. https://en.wikipedia.org/wiki/Political_polarization_in_the_United_States The IRA , social media and political polarization in the united states, 2012-2018 . Technical report, U.S. Sena...

    2019

  8. [16]

    Rohit Kumar Kaliyar, Anurag Goswami, and Pratik Narang. 2021. https://doi.org/10.1007/s11042-020-10183-2 Fakebert: Fake news detection in social media with a bert-based deep learning approach . Multim. Tools Appl., 80(8):11765--11788

    2021 doi

  9. [17]

    Andrea Landherr, Bettina Friedl, and Julia Heidemann. 2010. https://doi.org/10.1007/S12599-010-0127-3 A critical review of centrality measures in social networks . Bus. Inf. Syst. Eng., 2(6):371--385

    2010 doi

  10. [18]

    Jiawen Li, Yudianto Sujana, and Hung-Yu Kao. 2020. https://doi.org/10.18653/v1/2020.coling-main.473 Exploiting microblog conversation structures to detect rumors . In Proceedings of the 28th International Conference on Computational Linguistics, pages 5420--5429, Barcelona, Sp...

    2020 doi

  11. [19]

    Yang Liu and Yi - fang Brook Wu. 2018. https://doi.org/10.1609/AAAI.V32I1.11268 Early detection of fake news on social media through propagation path classification with recurrent and convolutional networks . In Proceedings of the Thirty-Second AAAI Conference on Artificial In...

    2018 doi

  12. [20]

    Sadek Rayhan, Mahdi Hasan Shuvo, Islam Sadia, and Md

    Fahim Belal Mahmud, Mahi Md. Sadek Rayhan, Mahdi Hasan Shuvo, Islam Sadia, and Md. Kishor Morol. 2022. https://arxiv.org/abs/2203.14132 A comparative analysis of graph neural networks and commonly used machine learning algorithms on fake news detection . CoRR, abs/2203.14132

    2022 arXiv

  13. [21]

    Rahul Mishra. 2020. https://doi.org/10.1109/CVPRW50498.2020.00334 Fake news detection using higher-order user to user mutual-attention progression in propagation paths . In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR Workshops 2020, Seattle, WA, U...

    2020

  14. [22]

    Bronstein

    Federico Monti, Fabrizio Frasca, Davide Eynard, Damon Mannion, and Michael M. Bronstein. 2019. https://arxiv.org/abs/1902.06673 Fake news detection on social media using geometric deep learning . CoRR, abs/1902.06673

    2019 arXiv

  15. [23]

    Massimo La Morgia, Alessandro Mei, and Alberto Maria Mongardini. 2023. https://arxiv.org/abs/2303.05345 Tgdataset: a collection of over one hundred thousand telegram channels . CoRR, abs/2303.05345

    2023 arXiv

  16. [24]

    Dan Saattrup Nielsen and Ryan McConville. 2022. https://doi.org/10.1145/3477495.3531744 Mumin: A large-scale multilingual multimodal fact-checked misinformation social network dataset . In SIGIR '22: The 45th International ACM SIGIR Conference on Research and Development in In...

    2022

  17. [25]

    Jeremy Nixon, Mike Dusenberry, Linchuan Zhang, Ghassen Jerfel, and Dustin Tran. 2019. https://arxiv.org/abs/1904.01685 Measuring calibration in deep learning . CoRR, abs/1904.01685

    2019 arXiv

  18. [26]

    Richard Rogers. 2020. https://doi.org/10.1177/0267323120922066 Deplatforming: Following extreme internet celebrities to telegram and alternative social media . European Journal of Communication, 35(3):213--229

    2020 doi

  19. [27]

    Kai Shu, Deepak Mahudeswaran, Suhang Wang, Dongwon Lee, and Huan Liu. 2020. https://doi.org/10.1089/big.2020.0062 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

  20. [28]

    Kai Shu, Amy Sliva, Suhang Wang, Jiliang Tang, and Huan Liu. 2017. https://doi.org/10.1145/3137597.3137600 Fake news detection on social media: A data mining perspective . SIGKDD Explor. Newsl., 19(1):22–36

    2017

  21. [29]

    Kai Shu, Xinyi Zhou, Suhang Wang, Reza Zafarani, and Huan Liu. 2019. https://doi.org/10.1145/3341161.3342927 The role of user profiles for fake news detection . In ASONAM '19: International Conference on Advances in Social Networks Analysis and Mining, Vancouver, British Colum...

    2019

  22. [30]

    Veronika Solopova, Viktoriia Herman, Christoph Benzm \"u ller, and Tim Landgraf. 2024. https://aclanthology.org/2024.eacl-demo.6 Check news in one click: NLP -empowered pro-kremlin propaganda detection . In Proceedings of the 18th Conference of the European Chapter of the Asso...

    2024

  23. [31]

    Abdullah All Tanvir, Ehesas Mia Mahir, S M Asiful Huda, and Shuvo Barua. 2020. https://doi.org/10.1109/AISP48273.2020.9073583 A hybrid approach for identifying authentic news using deep learning methods on popular twitter threads . In 2020 International Conference on Artificia...

    2020

  24. [32]

    Ramya Tekumalla and Juan M. Banda. 2023. https://doi.org/10.1007/978-3-031-48044-7_26 Leveraging large language models and weak supervision for social media data annotation: An evaluation using COVID-19 self-reported vaccination tweets . In HCI International 2023 - Late Breaki...

    2023 doi

  25. [33]

    Manasa, B

    Santosh Kumar Uppada, K. Manasa, B. Vidhathri, R. Harini, and B. Sivaselvan. 2022. https://doi.org/10.1007/S13278-022-00878-9 Novel approaches to fake news and fake account detection in osns: user social engagement and visual content centric model . Soc. Netw. Anal. Min., 12(1):52

    2022 doi

  26. [34]

    Aleksandra Urman and Stefan Katz. 2022. What they do in the shadows: examining the far-right networks on telegram. Information, communication & society, 25(7):904--923

    2022

  27. [35]

    Claire Wardle and Hossein Derakhshan. 2017. https://rm.coe.int/information-disorder-report-november-2017/1680764666 Information disorder: Toward an interdisciplinary framework for research and policymaking . Technical report, Council of Europe

    2017

  28. [36]

    Liang Wu and Huan Liu. 2018. https://doi.org/10.1145/3159652.3159677 Tracing fake-news footprints: Characterizing social media messages by how they propagate . In Proceedings of the Eleventh ACM International Conference on Web Search and Data Mining, WSDM 2018, Marina Del Rey,...

    2018

  29. [37]

    Xiaoyu Yang, Yuefei Lyu, Tian Tian, Yifei Liu, Yudong Liu, and Xi Zhang. 2020. https://doi.org/10.24963/ijcai.2020/197 Rumor detection on social media with graph structured adversarial learning . In Proceedings of the Twenty-Ninth International Joint Conference on Artificial I...

    2020 doi

  30. [38]

    Qiang Zhang, Aldo Lipani, Shangsong Liang, and Emine Yilmaz. 2019. https://doi.org/10.1145/3308558.3313718 Reply-aided detection of misinformation via bayesian deep learning . In The World Wide Web Conference, WWW 2019, San Francisco, CA, USA, May 13-17, 2019 , pages 2333--2343. ACM

    2019

  31. [39]

    Xitong Zhang, Yixuan He, Nathan Brugnone, Michael Perlmutter, and Matthew J. Hirn. 2021. https://proceedings.neurips.cc/paper/2021/hash/e32084632d369461572832e6582aac36-Abstract.html Magnet: A neural network for directed graphs . In Advances in Neural Information Processing Sy...

    2021

  32. [40]

    Xinyi Zhou, Jindi Wu, and Reza Zafarani. 2020. https://doi.org/10.1007/978-3-030-47436-2\_27 SAFE: similarity-aware multi-modal fake news detection . In Advances in Knowledge Discovery and Data Mining - 24th Pacific-Asia Conference, PAKDD 2020, Singapore, May 11-14, 2020, Proc...

    2020 doi

  33. [41]

    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...

  34. [42]

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