Pith. sign in

REVIEW 2 major objections 3 minor 49 references

Network Information Enhances Unreliable News Domain Detection

T0 review · 2 major / 3 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper establishes that statistically validated domain co-sharing networks systematically improve news-domain reliability classification, outperforming network-unaware baselines on identical features both with and without article text.

desk verdict A solid empirical case that network co-sharing improves domain reliability classification, but the test-domain evaluation protocol is underspecified — fix that before trusting 'systematically'. read the letter →

arxiv 2608.02399 v1 pith:4MC4653E submitted 2026-08-03 cs.SI cs.LGphysics.soc-ph

Raphaela Ke{ss}ler , Roman David Ventzke , Viola Priesemann , Giordano De Marzo This is my paper
classification cs.SIcs.LGphysics.soc-ph
keywords misinformationdetectionnewsdomainreliabilitygraphneuralnetworksTelegramco-sharingnetworkassortativemixingcontent-agnosticfeaturesbipartiteconfigurationmodel
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 tries to establish that the pattern of which news domains are shared together in Telegram chats carries reliable signal about those domains' trustworthiness, and that machine-learning models that exploit this network structure classify domain reliability better than models that see only the same content or spreading features without the graph. It builds a domain co-sharing network validated against a null model, shows it is assortative by reliability, and then compares Graph Neural Networks against a feature-only MLP. GraphSAGE achieves 0.63 accuracy with content features and 0.53 without them, versus 0.55 and 0.47 for the MLP. The practical stakes: content-based detection is increasingly unreliable as low-quality sources mimic journalism, so a content-agnostic network signal could still flag suspect domains even when text analysis is infeasible or misleading.

What carries the argument

The load-bearing object is the statistically validated domain co-sharing network, a monopartite graph of news domains constructed by projecting the chat–domain bipartite network and pruning edges with the Bipartite Configuration Model at significance p=0.01, so remaining edges indicate genuine co-sharing beyond what random degree sequences would produce. This network is what allows message-passing architectures to propagate reliability information from a domain's co-shared neighbours; the reliability assortativity of 0.22 is the mechanism they exploit. The specific architecture that benefits most is GraphSAGE, which samples and aggregates neighbourhood features, but the pattern holds across

What would settle it

Re-run the content-agnostic GraphSAGE evaluation with test domains detached from the graph (no edges at prediction time), or with a temporal split where test domains are not present in the training graph and their co-sharing edges are masked. If accuracy falls to the MLP baseline's 0.47, the gain comes from attaching test nodes to the graph; if it stays near 0.53, the relational structure generalises to previously unseen domains.

Watch

Extended reading notes

Core claim

The paper claims that a domain co-sharing network—where two news domains are connected if they are shared in the same Telegram chats more often than chance—carries a reliable signal about domain trustworthiness, and that graph-based classifiers can exploit it. The central discovery is that this network exhibits strong assortative mixing by reliability (assortativity coefficient 0.22, p<0.001) and that, on identical features, graph neural networks consistently beat a network-unaware MLP: GraphSAGE reaches 63% accuracy and F1 with content features (vs 55% for MLP) and 53% with only spreading dynamics (vs 47% for MLP). The gain is consistent across GCN, GAT, and GraphSAGE, and the paper interpr

Load-bearing premise

The result rests on test domains being attached to the validated co-sharing graph at inference with their edges to training domains available for message passing; the paper does not state this, so the 13–14% gain may depend on transductive access to the test domains' own co-sharing edges rather than on a generalisable relational rule.

Editorial extensions

If this is right

  • Reliability labels can be inferred from sharing structure alone: a content-agnostic GraphSAGE still reaches 0.53 accuracy, 0.06 above MLP, showing that co-sharing topology remains informative when text is absent.
  • Network-aware models are particularly valuable for catching unreliable domains: the MLP baseline in the content-agnostic setting never predicts the unreliable class, while all graph-aware models recover a substantial fraction (42–46 of 171 test domains).
  • Adding statistically validated network edges to any feature set yields consistent gains; the gap holds across architectures (GCN, GAT, GraphSAGE) and is largest for the inductive GraphSAGE, suggesting the benefit is structural rather than architecture-specific.
  • Domain-level reliability assessment can work on platforms like Telegram where content scraping is technically or legally difficult, as long as URL-sharing metadata is available.
  • The approach is language-independent in its content-agnostic variant, since it relies on spreading dynamics rather than text.

Reading between the lines

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

  • The authors do not specify whether test domains are attached to the validated graph at inference. A direct reading of the pipeline suggests they are, meaning the measured gain may be partly transductive; a cleaner test would detach test nodes or use a temporal split.
  • The Supplementary robustness check—denser networks from looser pruning thresholds do not improve performance—implies that the BiCM validation step itself is what concentrates the reliability signal. This suggests a transferable design rule: filter co-sharing edges by statistical significance before message passing on any platform.
  • If assortative mixing is the underlying mechanism, the method should transfer to other platforms where co-sharing is observable, and the expected gain should track the strength of reliability assortativity. Measuring both across platforms would be a direct test of the mechanism.
  • The paper discretises a continuous reliability score into three classes but cites best-practice guidance favouring continuous ratings. A graph-based regression on the continuous score could exploit the same network to predict fine-grained reliability, potentially extracting more signal from the assortative structure.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 3 minor

Summary. The paper asks whether network structure improves domain-level classification of unreliable news sources. Using Telegram URL-sharing data, the authors construct a chat–domain bipartite network, project it to a domain co-sharing network, and validate edges with the Bipartite Configuration Model. They show assortative mixing by an external reliability score (assortativity r=0.22) and then compare GNN architectures (GCN, GAT, GraphSAGE) against an MLP baseline on identical node features—both content embeddings from a multilingual SBERT model and content-agnostic spreading features. The main reported result is that GraphSAGE outperforms the MLP by 8 accuracy points with content features (0.63 vs 0.55) and by 6 points without content (0.53 vs 0.47). The authors conclude that network topology systematically improves reliability assessment and remains useful when content analysis is infeasible.

Significance. If the evaluation is sound, this is a worthwhile contribution to source-level misinformation detection: it operates at the domain level, uses an external fact-checking aggregation for labels, validates the co-sharing network with BiCM, compares models on identical features, and includes a transformer text baseline and robustness checks in the SI. The code is public. The main reason for caution is that the paper does not specify how the held-out test domains are placed in the graph at inference; the central claim about network information depends directly on this protocol, which the manuscript leaves ambiguous. The claim would be strengthened by repeated splits and error bars, given the modest margins for some architectures.

major comments (2)
  1. [Results/Evaluation; Methods Data Preprocessing, Tables 3–4] The central claim that network topology drives the GraphSAGE gains requires a precise statement of how the 1,222 test domains enter the graph. Table 3 lists the test domains as raw counts; Table 4 and the Results state that the validated network contains only the 4,574 training domains that survive BiCM validation. The manuscript never states whether test domains are attached to this graph at inference. If they are evaluated as isolated nodes, SAGEConv has no neighbors to aggregate and the model degenerates to a feature transform; the 0.63 vs 0.55 gap would then be an architecture/hyperparameter artifact rather than evidence for network information. If test domains are attached through co-sharing edges to training domains, the evaluation is transductive and the 'inductive learning'/'unseen nodes' statements in Models and Evaluation are unsupported. The paper must disclose the exact edge_
  2. [Results/Evaluation, Table 2; Hyperparameter Tuning] All results are point estimates from a single fixed 80/20 split, with no error bars, repeated runs, or significance tests. The 'systematically improves' claim rests on margins as small as 3 percentage points (GCN and GAT vs MLP in the content-agnostic setting), and the content-agnostic MLP never predicts the unreliable class (SI S5), so accuracy differences may be driven by class-level behavior. Please report mean and standard deviation over multiple random splits or bootstrap confidence intervals, and include per-class recall and F1, together with a statistical test of the GraphSAGE advantage. This is necessary to support the generalization claim made in the abstract and discussion.
minor comments (3)
  1. [Random and Baseline Model; Table 2] The 'Simple Network Classifier' baseline is not described in sufficient detail. The text refers to a neighbor-averaging baseline, but it is unclear how the average is computed for test domains, especially if test nodes are not in the validated graph. Define the procedure and state the graph used for this baseline.
  2. [Results, Figure 3 caption] The color mapping for node reliability is described as 'light blue' to 'dark', which is ambiguous. Specify the colormap and explicitly state which end corresponds to low versus high reliability in panel (b).
  3. [Methods, Data Preprocessing] Table 3's 'Validated' column is defined in the Feature Extraction subsection, but the table appears in Data Preprocessing. Add an explicit cross-reference or define the column where the table is first discussed.

Circularity Check

0 steps flagged · score 2.0 of 10

No circularity by construction; central comparison is an empirical test, with only a minor non-load-bearing self-citation and an unspecified inductive/transductive protocol that is a correctness risk.

full rationale

The paper's derivation chain is not circular. Labels come from an external fact-checking aggregation: 'we employ the comprehensive domain rating dataset compiled by Lin et al., which provides credibility assessments for 11,520 news domains [21]'; node features are text embeddings or spreading dynamics; and the graph is built from URL co-sharing in Telegram chats and pruned with the Bipartite Configuration Model, whose null model preserves degree sequences, not reliability. The reliability assortativity (r=0.22) is a measured property of that graph, not an input used to build it, so the subsequent GNN comparison is an empirical test. The GraphSAGE-vs-MLP comparison uses 'identical features' with the graph as the only difference, so the reported gain is not definitionally forced by a fitted parameter being renamed a prediction. Hyperparameters are tuned on a fixed validation split with test accuracy reported separately, so no fitted value is recycled as the headline result. The only self-citations are the authors' own Telegram dataset and its unguided-sample comparison ('We used a large Telegram data-set which has been obtained by crawling public groups chats and channels ... described in [25, 11]'); these are disclosed data-provenance and robustness checks, not a load-bearing theorem or ansatz imported to force the conclusion, and the target labels are external to that dataset. A genuine experimental gap exists: the paper never states whether the 1,222 test domains are attached to the 4,574-node validated graph at inference, so the claim that GraphSAGE generalizes to 'unseen nodes' is not fully supported and the measured gain could be transductive or an architecture artifact. That is a correctness/transparency issue, not circularity, because even in a transductive reading the graph edges are not derived from the target labels and no prediction reduces to its input by construction.

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

The central claim rests on external reliability labels, a network-construction assumption (co-sharing implies audience similarity), a statistical null model for edge validation, and an under-specified evaluation protocol. No new physical or conceptual entities are introduced.

free parameters (4)
  • PC1 class thresholds = 0.33 and 0.66
    Chosen by hand to discretize the continuous Lin et al. reliability score into unreliable/questionable/reliable; all models are evaluated on this fixed discretization, so it shapes the absolute accuracy and the practical meaning of the result, though it does not differentially advantage GNNs.
  • BiCM significance threshold = p=0.01
    Used to prune the monopartite projection; the authors report robustness to p=0.05 and p=0.5 in SI S3, so the central comparison is not highly sensitive to this choice.
  • Avalanche time window Δt = 1 hour
    Defines the bursty spreading features (avalanches, virality); chosen a priori and not tuned, but it affects all content-agnostic features and thus the content-agnostic GNN result.
  • Per-model hyperparameters = Table 6 (e.g., hidden size 64, dropout 0.3 for GraphSAGE content)
    Selected via grid search on validation accuracy; standard model selection, but the absence of repeated runs means the reported test accuracies are single draws from this selection procedure.
assumptions (4)
  • domain assumption The Lin et al. composite PC1 score is a valid and consistent ground-truth measure of news-domain reliability, and its three-bin discretization defines a meaningful classification task.
    All labels in the training/test split come from this external dataset (Results, Data Preprocessing); if the score is systematically biased for Telegram-popular domains, the measured GNN advantage may not transfer to true reliability.
  • domain assumption Co-sharing of domains in Telegram chats is a signal of audience/ideological proximity, not merely topic overlap or bot coordination.
    The entire network-based approach rests on this (Introduction, Results, Network Structure); the authors test assortativity against shuffled labels, but the underlying sociological assumption is not independently established in this paper.
  • domain assumption The Bipartite Configuration Model provides a valid null model whose p-values identify statistically significant co-occurrences.
    Used to prune edges (Methods Feature Extraction); while established in the literature, the correctness of the backbone depends on the null model's assumptions matching the data-generating process.
  • ad hoc to paper The GNN evaluation protocol (train/test graph membership) is well-defined and does not leak test labels into message passing.
    This is required for the reported GNN accuracies to be meaningful; Table 4 only lists training nodes, so the evaluation protocol is inferred rather than stated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Network Information Enhances Unreliable News Domain Detection." pith.science (2026). https://pith.science/paper/4MC4653E

@misc{pith2026260802399,
  author       = {Pith},
  title        = {Pith review of: Network Information Enhances Unreliable News Domain Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4MC4653E}},
  note         = {Machine review of arXiv:2608.02399}
}
read the original abstract

Content-based detection of unreliable news is increasingly difficult, as low-reliability sources mimic credible journalism and generative AI makes fabricated content harder to flag. We ask whether network structure can improve news reliability classification, taking a domain-level approach that shifts the focus from individual articles to source reliability. From URL-sharing patterns in Telegram chats, we build a statistically validated domain co-sharing network and find assortative mixing by reliability: low-reliability domains group together, as do reliable ones. Exploiting this structure, we compare Graph Neural Networks against network-unaware baselines using both content-aware features (multilingual text embeddings) and content-agnostic features (spreading dynamics). GNNs consistently outperform Multi-Layer Perceptrons on identical features, with GraphSAGE best in both settings (accuracy 0.63 with content, 0.53 without), a 13-14% relative gain over the network-unaware baseline. Network topology thus systematically improves domain reliability assessment, and remains effective even when content analysis is infeasible.

Figures

Figures reproduced from arXiv: 2608.02399 by the authors.

Figure 1
Figure 1. Schematic representation of the analysis. a) Each url is characterized by content features (text embedding) and content agnostic features (spreading features). We then average all urls’ features corresponding to the same domain to get content and spreading features at the domain level. b) The sharing of domains on Telegram can be represented by a bipartite network, with one layer containing the chats and the other t… view at source ↗
Figure 2
Figure 2. Distribution of news-domain reliability on Telegram. Histogram of the composite reliability score (PC1) of the news links shared across Telegram chats, obtained by aggregating six major fact-checking databases through principal component analysis; higher values denote more reliable sources. The dashed and dotted vertical lines mark the distribution mean (0.59) and median (0.63), while the two solid lines at 0.33 and… view at source ↗
Figure 3
Figure 3. Assortative mixing of reliability in the domain co-sharing network. a) For each domain, the average reliability score of its network neighbors is plotted against the domain’s own reliability score. The positive trend (blue regression line) reflects assortative mixing: domains tend to be co-shared with others of comparable reliability, quantified by a reliability assortativity coefficient of 0.22. b) Visualization of… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 7 linked inside Pith

  1. [1]

    Williams, and Kathleen M

    Peter Carragher, Evan M. Williams, and Kathleen M. Carley. Detection and Discovery of Misinformation Sources Using Attributed Webgraphs.Proceedings of the International AAAI Conference on Web and Social Media, 18:214–226, May 2024

    2024

  2. [2]

    Messaging apps: most popular by global downloads 2025, February 2025

    Laura Ceci. Messaging apps: most popular by global downloads 2025, February 2025

    2025

  3. [3]

    The echo chamber effect on social media.Proceedings of the National Academy of Sciences, 118(9):e2023301118, March 2021

    Matteo Cinelli, Gianmarco De Francisci Morales, Alessandro Galeazzi, Walter Quattrociocchi, and Michele Starnini. The echo chamber effect on social media.Proceedings of the National Academy of Sciences, 118(9):e2023301118, March 2021

    2021

  4. [4]

    The COVID-19 social media infodemic

    Matteo Cinelli, Walter Quattrociocchi, Alessandro Galeazzi, Carlo Michele Valensise, Emanuele Brugnoli, Ana Lucia Schmidt, Paola Zola, Fabiana Zollo, and Antonio Scala. The COVID-19 social media infodemic. Scientific Reports, 10(1):16598, October 2020

    2020

  5. [5]

    Collins 2017 Word of the Year Shortlist, November 2017

    Collins-Dictionary. Collins 2017 Word of the Year Shortlist, November 2017

    2017

  6. [6]

    langdetect: Port of google’s language-detection library to python

    Michal Danilák. langdetect: Port of google’s language-detection library to python. https://github.com/ Mimino666/langdetect, 2021

    2021

  7. [7]

    Du Rove’s Channel - Digital Resistance, 2018

    Pavel Durov. Du Rove’s Channel - Digital Resistance, 2018

    2018

  8. [8]

    Du Rove’s Channel - Telegram monthly active user, March 2025

    Pavel Durov. Du Rove’s Channel - Telegram monthly active user, March 2025

    2025

Show all 49 references
  1. [9]

    Fast graph representation learning with pytorch geometric.arXiv preprint arXiv:1903.02428, 2019

    Matthias Fey and Jan Eric Lenssen. Fast graph representation learning with pytorch geometric.arXiv preprint arXiv:1903.02428, 2019. 11 APREPRINT- AUGUST4, 2026

    1903 arXiv

  2. [10]

    Belief in Conspiracy Theories.Political Psychology, 15(4):731–742, 1994

    Ted Goertzel. Belief in Conspiracy Theories.Political Psychology, 15(4):731–742, 1994. Publisher: [International Society of Political Psychology, Wiley]

    1994

  3. [11]

    Mohr, Arne I

    Anastasia Golovin, Sebastian B. Mohr, Arne I. Gottwald, Ulrik Hvid, Srushhti Trivedi, Joao Pinheiro Neto, Andreas C. Schneider, and Viola Priesemann. Teragram: A structured longitudinal dataset of the telegram messenger.Proceedings of the International AAAI Conference on Web a...

    2026

  4. [12]

    Advancing Fake News Detection with Graph Neural Network and Deep Learning.Journal of Physics: Complexity, August 2024

    Haji Gul, Feras Al-Obeidat, Muhammad Wasim, Adnan Amin, and Fernando Moreira. Advancing Fake News Detection with Graph Neural Network and Deep Learning.Journal of Physics: Complexity, August 2024

    2024

  5. [13]

    Inductive Representation Learning on Large Graphs.Advances in Neural Information Processing Systems, 30, 2017

    Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive Representation Learning on Large Graphs.Advances in Neural Information Processing Systems, 30, 2017

    2017

  6. [14]

    Graph Neural Networks with Continual Learning for Fake News Detection from Social Media, August 2020

    Yi Han, Shanika Karunasekera, and Christopher Leckie. Graph Neural Networks with Continual Learning for Fake News Detection from Social Media, August 2020. arXiv:2007.03316 [cs]

    2020 arXiv

  7. [15]

    Aliaksandr Herasimenka, Jonathan Bright, Aleksi Knuutila, and Philip N. Howard. Misinformation and pro- fessional news on largely unmoderated platforms: the case of telegram.Journal of Information Technology & Politics, 20(2):198–212, April 2023

    2023

  8. [16]

    As- sociation of COVID-19 Misinformation with Face Mask Wearing and Social Distancing in a Nationally Rep- resentative US Sample.Health Communication, 36(1):6–14, January 2021

    Robert Hornik, Kikut , Ava, Jesch , Emma, Woko , Chioma, Siegel , Leeann, , and Kwanho Kim. As- sociation of COVID-19 Misinformation with Face Mask Wearing and Social Distancing in a Nationally Rep- resentative US Sample.Health Communication, 36(1):6–14, January 2021. Publishe...

    2021

  9. [17]

    sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 · Hugging Face, 2019

    Hugging-Face. sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 · Hugging Face, 2019

    2019

  10. [18]

    Fake news outbreak 2021: Can we stop the viral spread? Journal of Network and Computer Applications, 190:103112, September 2021

    Tanveer Khan, Antonis Michalas, and Adnan Akhunzada. Fake news outbreak 2021: Can we stop the viral spread? Journal of Network and Computer Applications, 190:103112, September 2021

    2021

  11. [19]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Semi-Supervised Classification with Graph Convolutional Networks, February

  12. [20]

    Thomas B Ksiazek, Limor Peer, and Kevin Lessard. User engagement with online news: Conceptualizing interactivity and exploring the relationship between online news videos and user comments.New Media & Society, 18(3):502–520, March 2016

    2016

  13. [21]

    High level of correspondence across different news domain quality rating sets.PNAS Nexus, 2(9):pgad286, September 2023

    Hause Lin, Jana Lasser, Stephan Lewandowsky, Rocky Cole, Andrew Gully, David G Rand, and Gordon Pennycook. High level of correspondence across different news domain quality rating sets.PNAS Nexus, 2(9):pgad286, September 2023

    2023

  14. [22]

    Best practices for source- based research on misinformation and news trustworthiness using newsguard.Journal of Quantitative Description: Digital Media, 5, January 2025

    Jula Lühring, Hannah Metzler, Ruggero Lazzaroni, Apeksha Shetty, and Jana Lasser. Best practices for source- based research on misinformation and news trustworthiness using newsguard.Journal of Quantitative Description: Digital Media, 5, January 2025

    2025

  15. [23]

    UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction, September 2020

    Leland McInnes, John Healy, and James Melville. UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction, September 2020. arXiv:1802.03426 [stat]

    2020 arXiv

  16. [24]

    A systematic literature review of the motivations to share fake news on social media platforms and how to fight them.New Media & Society, 26(2):1127–1150, February 2024

    Cristiane Melchior and Mírian Oliveira. A systematic literature review of the motivations to share fake news on social media platforms and how to fight them.New Media & Society, 26(2):1127–1150, February 2024

    2024

  17. [25]

    Inference of modular structures in dynamical systems and the application to telegram data

    Sebastian Bernd Mohr. Inference of modular structures in dynamical systems and the application to telegram data. Master’s thesis, Georg-August-Universität Göttingen, 2023

    2023

  18. [26]

    Bronstein

    Federico Monti, Fabrizio Frasca, Davide Eynard, Damon Mannion, and Michael M. Bronstein. Fake News Detection on Social Media using Geometric Deep Learning, February 2019. arXiv:1902.06673 [cs]

    2019 arXiv

  19. [27]

    MTEB: Massive Text Embedding Bench- mark, March 2023

    Niklas Muennighoff, Nouamane Tazi, Loïc Magne, and Nils Reimers. MTEB: Massive Text Embedding Bench- mark, March 2023. arXiv:2210.07316 [cs]

    2023 arXiv

  20. [28]

    Language detection library for java, 2010

    Shuyo Nakatani. Language detection library for java, 2010

    2010

  21. [29]

    Robertson, Amy Ross Arguedas, and Rasmus Kleis Nielsen

    Nic Newman, Richard Fletcher, Craig T. Robertson, Amy Ross Arguedas, and Rasmus Kleis Nielsen. Reuters Institute digital news report 2024. Technical report, Reuters Institute for the Study of Journalism, 2024

    2024

  22. [30]

    Thanh Thi Nguyen, Quoc Viet Hung Nguyen, Dung Tien Nguyen, Duc Thanh Nguyen, Thien Huynh-The, Saeid Nahavandi, Thanh Tam Nguyen, Quoc-Viet Pham, and Cuong M. Nguyen. Deep Learning for Deepfakes Creation and Detection: A Survey.arXiv preprint arXiv:1909.11573, 2019. Publisher: ...

    1909 arXiv

  23. [31]

    Universality, criticality and complexity of information propagation in social media.Nature Communications, 13(1), March 2022

    Daniele Notarmuzi, Claudio Castellano, Alessandro Flammini, Dario Mazzilli, and Filippo Radicchi. Universality, criticality and complexity of information propagation in social media.Nature Communications, 13(1), March 2022. 12 APREPRINT- AUGUST4, 2026

    2022

  24. [32]

    Word of the year, 2016

    Oxford-Dictionary. Word of the year, 2016

    2016

  25. [33]

    Pytorch: An imperative style, high-performance deep learning library.Advances in neural information processing systems, 32, 2019

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library.Advances in neural information processing systems, 32, 2019

    2019

  26. [34]

    Fake news detection: A survey of graph neural network methods.Applied Soft Computing, 139:110235, May 2023

    Huyen Trang Phan, Ngoc Thanh Nguyen, and Dosam Hwang. Fake news detection: A survey of graph neural network methods.Applied Soft Computing, 139:110235, May 2023

    2023

  27. [35]

    torch_geometric.nn.conv.GCNConv — pytorch_geometric documentation, 2025

    PyG-Team. torch_geometric.nn.conv.GCNConv — pytorch_geometric documentation, 2025

    2025

  28. [36]

    From virality to veracity: Examining false information on telegram vs

    Jonas Rieskamp, Milad Mirbabaie, Marie Langer, and Alexander Kocur. From virality to veracity: Examining false information on telegram vs. twitter. InProceedings of the 57th Hawaii International Conference on System Sciences, HICSS. Hawaii International Conference on System Sc...

    2024

  29. [37]

    Randomizing bipartite networks: the case of the world trade web.Scientific Reports, 5(1):10595, 2015

    Fabio Saracco, Riccardo Di Clemente, Andrea Gabrielli, and Tiziano Squartini. Randomizing bipartite networks: the case of the world trade web.Scientific Reports, 5(1):10595, 2015

    2015

  30. [38]

    Straka, Riccardo Di Clemente, Andrea Gabrielli, Guido Caldarelli, and Tiziano Squartini

    Fabio Saracco, Mika J. Straka, Riccardo Di Clemente, Andrea Gabrielli, Guido Caldarelli, and Tiziano Squartini. Inferring monopartite projections of bipartite networks: an entropy-based approach.New Journal of Physics, 19(5):053022, 2017

    2017

  31. [39]

    Mika J. Straka. Bipartite configuration model for python.https://github.com/tsakim/bicm, 2017

    2017

  32. [40]

    Fast and scalable likelihood maximization for Exponential Random Graph Models with local constraints.Scientific Reports, 11(1):15227, July 2021

    Nicolò Vallarano, Matteo Bruno, Emiliano Marchese, Giuseppe Trapani, Fabio Saracco, Giulio Cimini, Mario Zanon, and Tiziano Squartini. Fast and scalable likelihood maximization for Exponential Random Graph Models with local constraints.Scientific Reports, 11(1):15227, July 2021

    2021

  33. [41]

    30 Interesting Telegram Statistics you need to check (2025) - Skillademia, February 2023

    Layla Varela. 30 Interesting Telegram Statistics you need to check (2025) - Skillademia, February 2023. Section: Statistics

    2025

  34. [42]

    Graph Attention Networks, February 2018

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph Attention Networks, February 2018. arXiv:1710.10903

    2018 arXiv

  35. [43]

    Viral misinformation and echo chambers: the diffusion of rumors about genetically modified organisms on social media.Internet Research, 30(5):1547–1564, June 2020

    Xiaohui Wang and Yunya Song. Viral misinformation and echo chambers: the diffusion of rumors about genetically modified organisms on social media.Internet Research, 30(5):1547–1564, June 2020

    2020

  36. [44]

    Silvan Wehrli, Anna-Maria Hartner, T Sonia Boender, Bert Arnrich, and Christopher Irrgang. Information pathways and voids in critical german online communities during the covid-19 vaccination discourse: Cross- platform and mixed methods analysis.Journal of Medical Internet Res...

    2025

  37. [45]

    harbinger of freedom

    Mariëlle Wijermars, , and Tetyana Lokot. Is Telegram a “harbinger of freedom”? The performance, practices, and perception of platforms as political actors in authoritarian states.Post-Soviet Affairs, 38(1-2):125–145, March

  38. [46]

    Jiawei Zhang, Bowen Dong, and Philip S. Yu. FakeDetector: Effective Fake News Detection with Deep Diffusive Neural Network. In2020 IEEE 36th International Conference on Data Engineering (ICDE), pages 1826–1829, April 2020. ISSN: 2375-026X

    2020

  39. [47]

    Social Network Analysis: History, Concepts, and Research

    Mingxin Zhang. Social Network Analysis: History, Concepts, and Research. In Borko Furht, editor,Handbook of Social Network Technologies and Applications, pages 3–21. Springer US, New York, NY , 2010

    2010

  40. [48]

    Unreliable

    Xinyi Zhou and Reza Zafarani. Network-based fake news detection: A pattern-driven approach.ACM SIGKDD explorations newsletter, 21(2):48–60, 2019. 13 APREPRINT- AUGUST4, 2026 Supplementary Information S1 Dataset Description and Descriptive Statistics S1.1 Most Frequently Shared...

    2019

  41. [2022]

    Publisher: Routledge _eprint: https://doi.org/10.1080/1060586X.2022.2030645

    2022

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.