Pith. sign in

REVIEW 4 major objections 6 minor 60 references

CleanNews: a Network-aware Fake News Mitigation Architecture for Social Media

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read CleanNews fuses DeBERTa text embeddings with Node2Vec user-network embeddings so that a CNN + BiGRU/BiLSTM classifier can learn both linguistic and relational cues of misinformation, and then uses SparseShield network immunization to cut th

desk verdict A competent engineering paper whose claimed novelty—text+Node2Vec fusion—isn't supported by its own ablations; still worth a referee for the system integration and honest limitations. read the letter →

arxiv 2509.04489 v1 pith:L7QONQVB submitted 2025-09-01 cs.SI

Maria-Diana Cotelin , Ciprian-Octavian Truică , Elena-Simona Apostol This is my paper · ORCID
classification cs.SI
keywords fakenewsdetectionnetworkimmunizationembeddingfusionDeBERTaNode2Vecsocialmediamisinformationdeeplearningrumor
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

CleanNews is a three-stage system—preprocessing, detection, mitigation—for spotting fake news in social media posts. Its central bet is that a post's truthfulness depends not only on its wording but also on who wrote it and how users are connected, so it concatenates DeBERTa text embeddings with Node2Vec embeddings of the author's position in the propagation graph. The paper reports that this hybrid representation, fed into CNN-BiGRU and CNN-BiLSTM variants, reaches 0.758 accuracy on Twitter15 and 0.823 on Twitter16, competitive with but not clearly better than several state-of-the-art systems. In the mitigation stage, SparseShield immunization blocks roughly half of the nodes that would otherwise become active spreaders in simulated propagation, far outperforming random node removal.

What carries the argument

The key mechanism is the concatenated embedding vector formed by joining per-tweet DeBERTa word embeddings with per-author Node2Vec node embeddings, so that each post's representation carries both lexical meaning and the author's structural position in the user graph. This combined embedding is fed into a stack of CNN and bidirectional recurrent layers (BiLSTM or BiGRU), whose output drives a four-class classifier. A separate mitigation stage runs SparseShield or NetShield immunization on the graph, penalizing detected harmful nodes by halving their influence scores before selecting nodes to block.

What would settle it

Train the CleanNews detection module on Twitter15 and Twitter16 with the graph edges shuffled or replaced by random edges, keeping the text embeddings fixed. If accuracy on the held-out set stays within noise of the reported numbers, the network embedding is not carrying the relational signal the paper claims; alternatively, test on a dataset with no user graph at all and measure the drop.

Watch

Extended reading notes

Core claim

The central claim is that fusing textual and user-network embeddings at the input layer lets a relatively simple deep network (CNN plus bidirectional recurrent layers) learn both linguistic cues and relational context of misinformation, and that this is enough to build a practical real-time moderation pipeline. The architecture is modular: any word embedding and any node embedding can be swapped in. The paper's own experiments show that the fused embedding does not consistently beat text-only DeBERTa (on Twitter16, BiLSTM CNN with text-only scores 0.823 accuracy vs 0.793 with Node2Vec added), so the strongest load-bearing result is the mitigation component: SparseShield reduced active nodes

Load-bearing premise

The pipeline assumes the available graph is a meaningful user or propagation network, so that a Node2Vec embedding of each author's node carries relational signal useful for classification; if the graph is missing, noisy, or unrelated to how rumors spread, the fusion contribution collapses.

Editorial extensions

If this is right

  • If the fusion works, platforms can use the same text-plus-graph embedding both to flag false posts and to decide which accounts to restrict, integrating detection and intervention in one pipeline.
  • SparseShield's roughly 50% saved-node rate in simulation suggests that targeted immunization with only 5% of nodes blocked could meaningfully damp rumor cascades in real deployment.
  • Because the architecture is modular, swapping in a stronger language model or a richer node embedding should improve detection without retraining the rest of the pipeline.
  • The ablation results imply that text content, especially DeBERTa representations, carries most of the classification signal; the graph embedding contributes mainly to the mitigation stage.
  • The system already supports four rumor classes (true, false, unverified, non-rumor), so it can be evaluated against operational moderation tasks rather than only binary fake/real decisions.

Reading between the lines

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

  • The detection benefit of the fused embedding appears dataset-dependent: on Twitter16 the +Node2Vec variant actually lowered accuracy for the best text-only model, so the graph signal may matter more for mitigation than for classification.
  • A natural testable extension is to evaluate SparseShield on full-scale social graphs rather than the 5% subgraphs used here; savings may shrink or grow with network density and community structure.
  • The same modular pipeline could be retargeted to other harmful content, such as hate speech or coordinated inauthentic behavior, by swapping labels and rerunning immunization.
  • If graph embeddings are only weakly predictive on datasets without meaningful follower or interaction structure, the fusion's value will be limited to platforms that can actually harvest user-network data at scale.
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 CleanNews, a three-stage architecture for detecting and mitigating fake news on social media. The preprocessing stage cleans text and encodes it with DeBERTa; the detection stage combines CNN with BiLSTM/BiGRU layers and optionally concatenates Node2Vec user embeddings; the mitigation stage applies SparseShield, NetShield, or a random solver to immunize network nodes. Experiments on Twitter15 and Twitter16 report detection accuracies up to 0.758 and 0.823, with ablations over eight model variants, and mitigation simulations report SparseShield saving about half of the nodes from activation. The abstract and conclusions claim that the embedding fusion is a key contribution and that SparseShield is effective, but the evidence in the manuscript is mixed and internally inconsistent.

Significance. If the claims held, the paper would offer a useful modular and publicly available system combining text-based detection with network-aware intervention. The manuscript deserves credit for releasing code, performing systematic grid searches, evaluating preprocessing variants, and honestly listing limitations. However, the load-bearing evidence for the central contribution is not convincing: the Node2Vec fusion improves only some configurations and degrades the best text-only model on one dataset, the graph used for Node2Vec is never specified, and the mitigation comparison is undermined by inconsistent baselines. These issues make the current results insufficient to validate the architecture as presented, although they may be addressable with additional experiments and a revised presentation.

major comments (4)
  1. [§3.2.5, Algorithm 1 lines 10–12; Tables 6 and 13] The central claim that fusing DeBERTa word embeddings with Node2Vec user embeddings lets the model jointly learn linguistic and relational cues is not supported by the paper's own ablations. In Table 6, adding Node2Vec changes accuracy as follows: BiLSTM CNN 0.728→0.725, BiGRU CNN 0.745→0.742, CNN BiLSTM 0.695→0.688, CNN BiGRU 0.685→0.758. In Table 13: BiLSTM CNN 0.823→0.793, BiGRU CNN 0.768→0.781, CNN BiLSTM 0.805→0.817, CNN BiGRU 0.805→0.811. Thus only 4 of 8 configurations improve, and the best Twitter16 accuracy (0.823) is achieved by the text-only DeBERTa model. No repeated runs, standard deviations, or significance tests are reported, so the single +7.3 point gain on CNN BiGRU (Twitter15) cannot be distinguished from seed variance. The key contribution needs stronger empirical support or a revised, more cautious claim.
  2. [§4.1, §4.3.4, §4.4.4, Algorithm 1] The graph G used for Node2Vec is never defined. The manuscript does not state how nodes and edges are constructed from Twitter15/Twitter16, whether G is a follower graph, an interaction graph, or a propagation tree, nor its size, density, or preprocessing. Node2Vec embeddings depend entirely on this graph; without a precise specification, the experiments cannot be reproduced and the claimed relational signal is uninterpretable. Section 5.4 acknowledges that many datasets lack user network structure, but the experiments assume such a graph exists. The authors must specify how G was obtained for these datasets, including any random subgraph sampling used in the mitigation simulations.
  3. [§4.3.4, §4.4.4, Tables 9–11 and 15–17] The mitigation comparison is invalid because each algorithm is evaluated against a different unblocked baseline. For Twitter15, the unblocked number of activated nodes is 32 for SparseShield, 39 for Random Solver, and 39 for NetShield; for Twitter16, these are 23, 18, and 15 respectively. Since the baselines differ, 'saved nodes' cannot be compared across algorithms. There are also internal inconsistencies: e.g., the SparseShield Twitter15 text says 34 nodes become active without mitigation, while Table 9 reports 32; the NetShield Twitter16 text references Figure 10 instead of the correct figure. Repeated simulation runs with fixed seeds and a common baseline are needed to support the claim that SparseShield saves roughly half of the nodes.
  4. [§4.4.2, §4.4.3, §5.2, Conclusions; Tables 13–14] The paper contradicts itself about the best Twitter16 model. Section 4.4.2 states that 'BiLSTM CNN with DeBERTa embeddings scored the highest accuracy' and the Conclusions repeat that this model used 'only the DeBERTa embeddings' with accuracy 0.823. Yet Table 13 labels the 0.823 row as 'CleanNews +Node2Vec' and Table 14 lists 'CleanNews DeBERTa+Node2Vec' with 0.823. Additionally, Section 5.2 reports that cross-validation raised CNN BiGRU+Node2Vec accuracy from 0.758 to 0.775, but this improved number is not used in any of the summary tables or the comparison. These conflations make it unclear what the proposed system actually contributes on Twitter16 and what its true reported performance is.
minor comments (6)
  1. [Algorithm 1] Typographical issues: the parameter is spelled 'reccurent' twice, and the functions are spelled 'getWordEmbeding' and 'getNodeEmbeding'.
  2. [Table 5] The row 'CNN BiLSTM+N2V' appears twice; presumably one row should be 'CNN BiGRU+N2V'.
  3. [Figure 5] The caption says 'Confusion Matrix for CNN BiGRU+N2V', while the text in §4.3.2 refers to the model as 'CNN BiLSTM+N2V'. These should be reconciled.
  4. [Tables 7 and 14] The heading contains the typo 'sate-of-the-art' instead of 'state-of-the-art'.
  5. [§4.3.4 and §4.4.4] The descriptions of the simulator are underspecified: no seed, no number of runs, and no explanation of why the unblocked baselines differ across the three algorithms in the same setting (e.g., Twitter15 unblocked active nodes range from 32 to 39). Clarifying the simulation protocol would help.
  6. [Figure 17 and Table 17] The text for NetShield on Twitter16 refers to 'Figure 10' but should refer to Figure 17. The caption also says 'NetShield Solver' inconsistently.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CleanNews' empirical pipeline is self-contained, and its self-cited components (SparseShield, GETAE, DANES) are used as algorithms or baselines rather than as load-bearing justifications.

full rationale

CleanNews is an empirical systems paper, not a derivation. The central claims—that DeBERTa+Node2Vec fusion improves detection and that SparseShield mitigates spread—are supported by the paper's own ablation studies, hyperparameter searches, and simulations. Algorithm 1 defines the architecture directly; the concatenation of word and node embeddings (line 12) is a supervised feature-engineering choice, not a fitted parameter renamed as a prediction. The reported accuracies come from training and evaluating classifiers on Twitter15/Twitter16, and the mitigation numbers come from an in-paper simulator. The paper's self-citations to SparseShield [3], GETAE [50], and DANES [59] are used as prior algorithms or comparison baselines; none is invoked as an unexamined premise that forces the paper's results. The limitations section explicitly concedes weak points (small datasets, lack of network structure, performance 'similar to state-of-the-art'), and no uniqueness theorem or ansatz is imported from the authors' prior work. While the novelty of the embedding fusion is debatable given GETAE/DANES, that is a matter of contribution framing, not circularity. The failure of Node2Vec to consistently help in ablations is a correctness/robustness concern, not a circular reduction.

Assumptions & free parameters 10 free parameters · 5 assumptions · 0 invented entities

The core experimental claim rests on standard transfer-learning assumptions, a graph-construction assumption that the authors themselves flag as often unmet, and a simulation model that is not validated against real spread data. In addition, a large number of hyperparameters and simulation choices are fit to or selected on the evaluation data, so the reported accuracies are best-case configurations rather than parameter-free predictions.

free parameters (10)
  • hidden size = 256 or 512, selected per model (Table 5/12)
    Grid-searched on each dataset; changes model capacity.
  • num layers = 3 or 4
    Grid-searched; depth of recurrent stack.
  • dropout = 0.1 or 0.2
    Grid-searched regularization.
  • learning rate = 0.0001 or 0.0003
    Grid-searched optimizer step.
  • num epochs = 30 or 40
    Grid-searched training length.
  • batch size = 24
    Fixed in grid.
  • immunization budget k = 5% of nodes, 134 (Twitter15) and 66 (Twitter16)
    Chosen by hand as 5% of random subgraph.
  • harmful node score penalty = 0.5 multiplier
    Equation (1) arbitrarily halves scores of harmful nodes.
  • random subgraph selection = 5% of nodes, no seed reported
    Determines which nodes and how many harmful nodes are in the simulation; differs across algorithms.
  • Node2Vec embedding dimension = not reported
    Dimension s of node embeddings is used in Algorithm 1 but never specified.
assumptions (5)
  • domain assumption Twitter15/Twitter16 labels are reliable for evaluating rumor detection
    All detection metrics depend on these labels; the authors note lack of expert annotation.
  • domain assumption A propagation graph can be constructed from the datasets and Node2Vec embeddings capture relational cues
    Algorithm 1 line 11 and Section 3.2.5 use graph G to train node embeddings; the paper's own limitation section says such network structure is often absent.
  • domain assumption The two-iteration active-node simulation approximates real misinformation spread
    Sections 4.3.4 and 4.4.4 evaluate mitigation solely through this simulator.
  • domain assumption Pretrained DeBERTa and Node2Vec representations transfer to this domain
    Used as fixed embeddings in all models without task-specific fine-tuning.
  • domain assumption CleanNews (4-class) and GETAE (binary) results are directly comparable
    Tables 7 and 14 place GETAE's binary metrics next to CleanNews' 4-class metrics with no adjustment or caveat.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CleanNews: a Network-aware Fake News Mitigation Architecture for Social Media." pith.science (2026). https://pith.science/paper/L7QONQVB

@misc{pith2026250904489,
  author       = {Pith},
  title        = {Pith review of: CleanNews: a Network-aware Fake News Mitigation Architecture for Social Media},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L7QONQVB}},
  note         = {Machine review of arXiv:2509.04489}
}
read the original abstract

With the widespread use of the internet and handheld devices, social media now holds a power similar to that of old newspapers. People use social media platforms for quick and accessible information. However, this convenience comes with a variety of risks. Anyone can freely post content, true or false, with the probability of remaining online forever. This makes it crucial to identify and tackle misinformation and disinformation on online platforms. In this article, we propose CleanNews, a comprehensive architecture to identify fake news in real-time accurately. CleanNews uses advanced deep learning architectures, combining convolutional and bidirectional recurrent neural networks, i.e., LSTM and GRU, layers to detect fake news. A key contribution of our work is a novel embedding technique that fuses textual information with user network structure, allowing the model to jointly learn linguistic and relational cues associated with misinformation. Furthermore, we use two network immunization algorithms, i.e., SparseShield and NetShield, to mitigate the spread of false information within networks. We conduct extensive ablation studies to evaluate the contribution of each model component and systematically tune hyperparameters to maximize performance. The experimental evaluation on two real-world datasets shows the efficacy of CleanNews in combating the spread of fake news.

Figures

Figures reproduced from arXiv: 2509.04489 by the authors.

Figure 1
Figure 1. CleanNews architecture datasets. There are a limited number of such datasets, and they are often in English and focused on political news. This affects the development and evaluation of detection models. The lack of large-scale, diverse, and real-world datasets is a significant challenge [42]. While the use of LLMs to generate contextually relevant synthetic rumor instances addresses the class imbalance problem, ver… view at source ↗
Figure 2
Figure 2. CleanNews Detection Architecture The posts are classified as true, false, non-rumor, or unverified based on learned patterns and features. By combining Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory networks (BiLSTM) / Bidirectional Gated Recurrent Unit (BiGRU), the models capture both local and global dependencies in the text. This allows them to better understand subtle characteristic… view at source ↗
Figure 3
Figure 3. Fake news detection result [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Fake news mitigation 7 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Confusion Matrix for CNN BiGRU+N2V profiles. GCAN and GCAN-G [32] incorporate user features into retweet propagation modeling, where GCAN leverages both recurrent and convolutional components, while GCAN￾G omits the convolutional network. SSGE [58] employs graph embedd…
Figure 8
Figure 8. Figure 8: Simulation of Information Spread using Random Solver for Twitter15 [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: shows that the blocked nodes are not necessarily close to the harmful [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Simulation of Information Spread using NetShield for Twitter15 [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Nodes after NetShield for Twitter15 4.4. Twitter16 4.4.1. Hyperparameters tuning For this dataset, we used the same Grid Search parameters as for the Twitter15 dataset ( [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 13
Figure 13. Figure 13: Simulation of Information Spread using SparseShield for Twitter16 [PITH_FULL_IMAGE:figures/full_fig_p014_13.png]
Figure 14
Figure 14. Figure 14: Nodes after SparseShield for Twitter16 Random Solver. Without a mitigation strategy, on average, 18 nodes are active by the end of the simulation, 8 nodes are active initially, and after another iteration, 18 nodes. With a Random blocking strategy, we obtain the same …
Figure 16
Figure 16. Figure 16: Nodes after Random Solver for Twitter15 [PITH_FULL_IMAGE:figures/full_fig_p015_16.png]
Figure 17
Figure 17. Figure 17: Simulation of Information Spread using NetShield Solver for [PITH_FULL_IMAGE:figures/full_fig_p015_17.png]
Figure 18
Figure 18. Figure 18: Nodes after NetShield for Twitter16 5. Discussions and Limitations In this section, the results are highlighted and discussed. 5.1. Preprocessing Module During this step, we experimented with classic preprocessing, aggressive preprocessing, soft preprocessing, and no …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

60 extracted references · 52 canonical work pages

  1. [1]

    Sherstoboeva, Russian Bans on ‘Fake News’ about the war in Ukraine: Conditional truth and unconditional loyalty, International Communication Gazette 86 (2024) 36–54

    E. Sherstoboeva, Russian Bans on ‘Fake News’ about the war in Ukraine: Conditional truth and unconditional loyalty, International Communication Gazette 86 (2024) 36–54. 17

    work page 2024

  2. [2]

    S. P. Reis, H. S. Wald, The Hamas massacre of Oct 7, 2023, and its aftermath, medical crimes, and the Lancet commission report on medicine, Nazism, and the Holocaust, Israel Journal of Health Policy Research 13 (2024) 19(1–7). doi:10.1186/s13584-024-00608-w

    work page doi:10.1186/s13584-024-00608-w 2023

  3. [3]

    Petrescu, C.-O

    A. Petrescu, C.-O. Truic ˘a, E.-S. Apostol, P. Karras, SparseShield: Social Network Immunization vs. Harmful Speech, in: ACM Conference on Information and Knowledge Management, 2021, pp. 1426–1436. doi:10. 1145/3459637.3482481

    work page arXiv 2021

  4. [4]

    C. Chen, H. Tong, B. A. Prakash, C. E. Tsourakakis, T. Eliassi-Rad, C. Faloutsos, D. H. Chau, Node Immunization on Large Graphs: Theory and Algorithms, IEEE Transactions on Knowledge and Data Engineering 28 (2015) 113–126. doi:10.1109/TKDE.2015.2465378

    work page arXiv 2015

  5. [5]

    Ilie, C.-O

    V .-I. Ilie, C.-O. Truic ˘a, E.-S. Apostol, A. Paschke, Context-Aware Misinformation Detection: A Benchmark of Deep Learning Architectures Using Word Embeddings, IEEE Access 9 (2021) 162122–162146. doi:10.1109/ACCESS.2021.3132502

    work page arXiv 2021

  6. [6]

    Mallik, S

    A. Mallik, S. Kumar, Word2Vec and LSTM based deep learning technique for context-free fake news detection, Multimedia Tools and Applications 83 (2024) 919–940

    work page 2024

  7. [7]

    Taher, A

    Y . Taher, A. Moussaoui, F. Moussaoui, Automatic fake news detection based on deep learning, FasTtext and news title, International Journal of Advanced Computer Science and Applications 13 (2022)

    work page 2022

  8. [8]

    Abualigah, Y

    L. Abualigah, Y . Y . Al-Ajlouni, M. S. Daoud, M. Altalhi, H. Migdady, Fake news detection using recurrent neural network based on bidirectional LSTM and GloVe, Social Network Analysis and Mining 14 (2024) 40

    work page 2024

Show all 60 references
  1. [9]

    Cotelin, E.-S

    M.-D. Cotelin, E.-S. Apostol, C.-O. Truic ˘a, NetGuardAI at EXIST2025: Sexism Detection using mDeBERTa, in: Working Notes of the Conference and Labs of the Evaluation Forum, 2025, pp. 1889–1897

    2025

  2. [10]

    Petrescu, C.-O

    A. Petrescu, C.-O. Truic ˘a, E.-S. Apostol, Language-based Mixture of Transformers for EXIST2024, in: Working Notes of the Conference and Labs of the Evaluation Forum, volume 3740 of CEUR Workshop Proceedings, 2024, pp. 1157–1164

    2024

  3. [11]

    Truic ˘a, E.-S

    C.-O. Truic ˘a, E.-S. Apostol, MisRoBÆRTa: Transformers versus Misinformation, Mathematics 10 (2022) 1–25(569). doi: 10.3390/ math10040569

    2022

  4. [12]

    Petrescu, C.-O

    A. Petrescu, C.-O. Truic ˘a, E.-S. Apostol, Language-based Mixture of Transformers for Sexism Identification in Social Networks, in: Conference and Labs of the Evaluation Forum, 2025

    2025

  5. [13]

    Truic ˘a, E.-S

    C.-O. Truic ˘a, E.-S. Apostol, A. Paschke, Awakened at CheckThat! 2022: Fake News Detection using BiLSTM and sentence transformer, in: Working Notes of the Conference and Labs of the Evaluation Forum, 2022, pp. 749–757

    2022

  6. [14]

    Truic ˘a, E.-S

    C.-O. Truic ˘a, E.-S. Apostol, It’s all in the Embedding! Fake News Detection using Document Embeddings, Mathematics 11 (2023) 1– 29(508). doi:10.3390/math11030508

    2023 doi

  7. [15]

    Badjatiya, S

    P. Badjatiya, S. Gupta, M. Gupta, V . Varma, Deep learning for hate speech detection in tweets, in: International Conference on World Wide Web, 2017, pp. 759–760

    2017

  8. [16]

    S. Raza, C. Ding, Fake news detection based on news content and social contexts: a transformer-based approach, International Journal of Data Science and Analytics 13 (2022) 335–362

    2022

  9. [17]

    H. N. Tran, U. Kruschwitz, ur-iw-hnt at CheckThat!-2022: Cross-lingual Text Summarization for Fake News Detection, in: CLEF (Working Notes), 2022, pp. 740–748

    2022

  10. [18]

    Farhangian, R

    F. Farhangian, R. M. Cruz, G. D. Cavalcanti, Fake news detection: Taxonomy and comparative study, Information Fusion 103 (2024) 102140

    2024

  11. [19]

    Szczepa ´nski, M

    M. Szczepa ´nski, M. Pawlicki, R. Kozik, M. Chora ´s, New explainability method for BERT-based model in fake news detection, Scientific reports 11 (2021) 23705

    2021

  12. [20]

    R. K. Kaliyar, A. Goswami, P. Narang, FakeBERT: Fake news detection in social media with a BERT-based deep learning approach, Multimedia tools and applications 80 (2021) 11765–11788

    2021

  13. [21]

    Dhiman, A

    P. Dhiman, A. Kaur, D. Gupta, S. Juneja, A. Nauman, G. Muhammad, GBERT: A hybrid deep learning model based on GPT-BERT for fake news detection, Heliyon 10 (2024)

    2024

  14. [22]

    B. Hu, Q. Sheng, J. Cao, Y . Shi, Y . Li, D. Wang, P. Qi, Bad actor, good advisor: Exploring the role of large language models in fake news detection, in: AAAI Conference on Artificial Intelligence, 2024, pp. 22105–22113

    2024

  15. [23]

    Anirudh, M

    K. Anirudh, M. Srikanth, A. Shahina, Multilingual fake news detection in low-resource languages: A comparative study using BERT and GPT-3.5, in: International Conference on Speech and Language Technologies for Low-resource Languages, Springer, 2023, pp. 387–397

    2023

  16. [24]

    L. S. Anjos, S. E. Quincozes, J. F. Kazienko, V . E. Quincozes, Investigating the performance of the GPT-3.5 model in fake news detection: An experimental analysis, in: Simp ´osio Brasileiro de Seguranc ¸a da Informac ¸˜ao e de Sistemas Computacionais (SBSeg), SBC, 2023, pp. 552–557

    2023

  17. [25]

    K. I. Roumeliotis, N. D. Tselikas, D. K. Nasiopoulos, Fake News Detection and Classification: A Comparative Study of Convolutional Neural Networks, Large Language Models, and Natural Language Processing Models, Future Internet 17 (2025)

    2025

  18. [26]

    J. Wu, J. Guo, B. Hooi, Fake News in Sheep’s Clothing: Robust Fake News Detection Against LLM-Empowered Style Attacks, in: ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 3367–3378

    2024

  19. [27]

    Askarizade, Enhancing rumor detection with data augmentation and generative pre-trained transformer, Expert Systems with Applications 262 (2025) 125649

    M. Askarizade, Enhancing rumor detection with data augmentation and generative pre-trained transformer, Expert Systems with Applications 262 (2025) 125649

    2025

  20. [28]

    Zhang, Q

    Z. Zhang, Q. Lv, X. Jia, W. Yun, G. Miao, Z. Mao, G. Wu, GBCA: Graph Convolution Network and BERT combined with Co-Attention for fake news detection, Pattern Recognition Letters 180 (2024) 26–32

    2024

  21. [29]

    T. Liu, Q. Cai, C. Xu, B. Hong, F. Ni, Y . Qiao, T. Yang, Rumor Detection with A Novel Graph Neural Network Approach, Academic Journal of Science and Technology 10 (2024) 305–310. doi:10.54097/farmdr42

    2024 doi

  22. [30]

    P. Meel, D. K. Vishwakarma, Fake news detection using semi-supervised graph convolutional network, arXiv preprint arXiv:2109.13476 (2021)

    2021 arXiv

  23. [31]

    Y . Wang, S. Qian, J. Hu, Q. Fang, C. Xu, Fake news detection via knowledge-driven multimodal graph convolutional networks, in: International Conference on Multimedia Retrieval, 2020, pp. 540–547

    2020

  24. [32]

    Lu, C.-T

    Y .-J. Lu, C.-T. Li, Gcan: Graph-aware co-attention networks for explainable fake news detection on social media, in: Annual Meeting of the Association for Computational Linguistics, ACL, 2020, pp. 505–514. doi:10.18653/v1/2020.acl-main.48

    2020 doi

  25. [33]

    S. Qian, J. Hu, Q. Fang, C. Xu, Knowledge-aware multi-modal adaptive graph convolutional networks for fake news detection, ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM) 17 (2021) 1–23

    2021

  26. [34]

    Ahammad, A

    M. Ahammad, A. Sani, K. Rahman, M. T. Islam, M. M. R. Masud, M. M. Hassan, M. A. T. Rony, S. M. N. Alam, M. S. H. Mukta, RoBERTa-GCN: A Novel Approach for Combating Fake News in Bangla Using Advanced Language Processing and Graph Convolutional Networks, IEEE Access 12 (2024) 1...

    2024

  27. [35]

    Petrescu, C.-O

    A. Petrescu, C.-O. Truic ˘a, E.-S. Apostol, Sentiment Analysis of Events in Social Media, in: International Conference on Intelligent Computer Communication and Processing, IEEE, 2019, pp. 143–149. doi: 10. 1109/iccp48234.2019.8959677

    2019

  28. [36]

    Petrescu, C.-O

    A. Petrescu, C.-O. Truic ˘a, E.-S. Apostol, A. Paschke, EDSA-Ensemble: an Event Detection Sentiment Analysis Ensemble Architecture, IEEE Transactions on Affective Computing 16 (2025) 555–572. doi:10.1109/ TAFFC.2024.3434355

    2025

  29. [37]

    Truic ˘a, E.-S

    C.-O. Truic ˘a, E.-S. Apostol, T. S , tefu, P. Karras, A Deep Learning Architecture for Audience Interest Prediction of News Topic on Social Media, in: International Conference on Extending Database Technology (EDBT2021), 2021, pp. 588–599. doi:10.5441/002/EDBT.2021.69

    2021 doi

  30. [38]

    Apostol, ¨Ozgur Coban, C.-O

    E.-S. Apostol, ¨Ozgur Coban, C.-O. Truic ˘a, CONTAIN: A community- based algorithm for network immunization, Engineering Science and Technology, an International Journal 55 (2024) 1–10(101728). doi: 10. 1016/j.jestch.2024.101728

    2024

  31. [39]

    Truic ˘a, E.-S

    C.-O. Truic ˘a, E.-S. Apostol, R.-C. Nicolescu, P. Karras, MCWDST: A Minimum-Cost Weighted Directed Spanning Tree Algorithm for Real- Time Fake News Mitigation in Social Media, IEEE Access 11 (2023) 125861–125873. doi:10.1109/ACCESS.2023.3331220

    2023

  32. [40]

    Truic ˘a, A.-T

    C.-O. Truic ˘a, A.-T. Constantinescu, E.-S. Apostol, StopHC: A Harmful Content Detection and Mitigation Architecture for Social Media Platforms, in: IEEE International Conference on Intelligent Computer Communication and Processing, 2024, pp. 1–5. doi:10.1109/ ICCP63557.2024.10793051

    2024

  33. [41]

    Apostol, C.-O

    E.-S. Apostol, C.-O. Truic ˘a, A. Paschke, ContCommRTD: A Distributed Content-Based Misinformation-Aware Community Detection System for Real-Time Disaster Reporting, IEEE Transactions on Knowledge and 18 Data Engineering (2024) 1–12. doi:10.1109/tkde.2024.3417232

    2024

  34. [42]

    B. Hu, Z. Mao, Y . Zhang, An overview of fake news detection: From a new perspective, Fundamental Research 5 (2025) 332–346. doi: 10. 1016/j.fmre.2024.01.017

    2025

  35. [43]

    Capuano, G

    N. Capuano, G. Fenza, V . Loia, F. D. Nota, Content-based fake news detection with machine and deep learning: A systematic review, Neurocomputing 530 (2023) 91–103

    2023

  36. [44]

    P. He, X. Liu, J. Gao, W. Chen, DeBERTa: Decoding-enhanced BERT with Disentangled Attention, in: International Conference on Learning Representations, 2021

    2021

  37. [45]

    Grover, J

    A. Grover, J. Leskovec, node2vec: Scalable Feature Learning for Networks, in: ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, ACM, 2016, pp. 855–864. doi: 10.1145/ 2939672.2939754

    2016

  38. [46]

    Truic ˘a, J

    C.-O. Truic ˘a, J. Darmont, J. Velcin, A Scalable Document-Based Architecture for Text Analysis, in: International Conference on Advanced Data Mining and Applications, Springer, 2016, pp. 481–494. doi: 10. 1007/978-3-319-49586-6_33

    2016

  39. [47]

    Truic ˘a, A

    C.-O. Truic ˘a, A. Guille, M. Gauthier, CATS: Collection and Analysis of Tweets Made Simple, in: ACM Conference on Computer Supported Cooperative Work and Social Computing Companion, ACM, 2016, pp. 41–44. doi:10.1145/2818052.2874320

    2016

  40. [48]

    J. Ma, W. Gao, K.-F. Wong, Detect Rumors in Microblog Posts Using Propagation Structure via Kernel Learning, in: Annual Meeting of the Association for Computational Linguistics, ACL, 2017, pp. 708–717. doi:10.18653/v1/P17-1066

    2017 doi

  41. [49]

    Truic ˘a, C

    C.-O. Truic ˘a, C. A. Leordeanu, Classification of an imbalanced data set using decision tree algorithms, Univiversity Politechnica of Bucharest Scientific Bulletin - Series C Electrical Engineering and Computer Science 79 (2017) 69–84

    2017

  42. [50]

    Truic ˘a, E.-S

    C.-O. Truic ˘a, E.-S. Apostol, M. Marogel, A. Paschke, GETAE: Graph Information Enhanced Deep Neural NeTwork Ensemble ArchitecturE for fake news detection, Expert Systems with Applications 275 (2025) 126984. doi:10.1016/j.eswa.2025.126984

    2025

  43. [51]

    Castillo, M

    C. Castillo, M. Mendoza, B. Poblete, Information credibility on twitter, in: Proceedings of the 20th international conference on World wide web, ACM, 2011, pp. 675–684. doi:10.1145/1963405.1963500

    2011

  44. [52]

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

    2016

  45. [53]

    S. Kwon, M. Cha, K. Jung, Rumor detection over varying time windows, PLOS ONE 12 (2017) 1–19(e0168344). doi: 10.1371/journal.pone. 0168344

    2017 doi

  46. [54]

    Y . Yang, L. Zheng, J. Zhang, Q. Cui, Z. Li, P. S. Yu, Ti-cnn: Convolutional neural networks for fake news detection, 2018. doi: 10.48550/ARXIV. 1806.00749

    2018 doi arXiv

  47. [55]

    Liu, Y .-F

    Y . Liu, Y .-F. Wu, Early detection of fake news on social media through propagation path classification with recurrent and convolutional networks, in: Proceedings of the AAAI Conference on Artificial Intelligence, AAAI, 2018, pp. 355–361. doi: 10.1609/aaai.v32i1. 11268

    2018 doi

  48. [56]

    Ruchansky, S

    N. Ruchansky, S. Seo, Y . Liu, Csi: A hybrid deep model for fake news detection, in: Proceedings of the 2017 ACM on Conference on Information and Knowledge Management, ACM, 2017, pp. 797–806. doi:10.1145/3132847.3132877

    2017

  49. [57]

    K. Shu, L. Cui, S. Wang, D. Lee, H. Liu, defend: Explainable fake news detection, in: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, ACM, 2019, pp. 395–405. doi:10.1145/3292500.3330935

    2019

  50. [58]

    D. T. Vu, J. J. Jung, Rumor detection by propagation embedding based on graph convolutional network, International Journal of Computational Intelligence Systems 14 (2021) 1053–1065. doi: 10.2991/ijcis.d. 210304.002

    2021 doi

  51. [59]

    Truic ˘a, E.-S

    C.-O. Truic ˘a, E.-S. Apostol, P. Karras, DANES: Deep Neural Network Ensemble Architecture for Social and Textual Context-aware Fake News Detection, Knowledge-Based Systems 294 (2024) 1–13(111715). doi:10.1016/j.knosys.2024.111715

    2024

  52. [60]

    Pennington, R

    J. Pennington, R. Socher, C. D. Manning, GloVe: Global vectors for word representation, in: Conference on Empirical Methods in Natural Language Processing, 2014, pp. 1532–1543. doi: 10.3115/v1/ D14-1162. 19

    2014 doi

Pith tools

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