CleanNews: Network-Aware Fake News Mitigation
- CleanNews is a network-aware fake news mitigation architecture that integrates hybrid deep neural networks with user network structure for real-time misinformation detection and control.
- The detection stage combines text embeddings (DeBERTa) with user network embeddings (Node2Vec) via CNN-BiGRU or CNN-BiLSTM models to capture both local and global dependencies.
- The mitigation stage employs network immunization algorithms like SparseShield and NetShield to strategically immunize key nodes, reducing misinformation spread by up to 50%.
CleanNews is a network-aware fake news mitigation architecture for social media that couples real-time detection with active containment of misinformation spread. It is defined by three linked elements: a preprocessing stage, a detection stage built from hybrid deep neural architectures, and a mitigation stage based on network immunization. Its central technical claim is that fake-news identification improves when textual information is fused with user network structure, and that post-detection intervention can be operationalized through SparseShield and NetShield to limit downstream diffusion (Cotelin et al., 1 Sep 2025).
1. System concept and architectural scope
CleanNews is presented as a modular, end-to-end architecture that simultaneously detects fake news and limits its dissemination in social media environments. Its distinguishing feature is the incorporation of both content and social network attributes in model training and intervention. The overall workflow is split into three interconnected modules: preprocessing, detection, and mitigation. The system also supports real-time detection and mitigation and integrates a user-friendly graphical interface for practical deployment (Cotelin et al., 1 Sep 2025).
| Module | Stated role |
|---|---|
| Preprocessing | Cleans and encodes data for downstream analysis |
| Detection | Employs deep neural architectures, leveraging both text and user network structure |
| Mitigation | Utilizes advanced network immunization algorithms to contain identified misinformation |
Within this design, CleanNews is not only a classifier. The mitigation module is part of the core architecture rather than an external add-on. This distinguishes the system conceptually from pipelines that terminate at article scoring or ranking. A related point, explicitly stated in the paper, is that the architecture is intended for real-time use rather than offline corpus analysis alone (Cotelin et al., 1 Sep 2025).
2. Detection module and hybrid neural architectures
The detection component uses hybrid deep neural networks that combine convolutional neural networks with bidirectional recurrent neural networks. The recurrent alternatives are BiLSTM and BiGRU, and the supported templates are BiRNN_CNN and CNN_BiRNN. The architecture flow is described as follows: an embedding layer accepts concatenated text and node embeddings; an optional BiRNN layer models sequential context; CNN layers with ReLU and MaxPooling extract local and hierarchical feature patterns; an optional BiRNN layer is used depending on the template; fully connected layers map features to output classes with Sigmoid or Softmax; and Dropout regularizes the network (Cotelin et al., 1 Sep 2025).
The functional division of labor is explicit. CNN layers are used to capture local spatial or textual patterns such as n-grams and phrase structure and are described as efficient for short posts such as tweets and comments. BiLSTM and BiGRU process sequences in forward and backward directions, capturing dependencies from both preceding and succeeding context. The paper states that the hybrid combination captures both local and global dependencies and thereby supports nuanced linguistic analysis for fake news detection (Cotelin et al., 1 Sep 2025).
The detection stage is therefore not a generic text classifier. It is a family of architectures whose layout can place convolutional extraction before or after bidirectional sequence modeling. This suggests that CleanNews treats rumor classification as a joint problem of local lexical regularities and longer contextual dependencies rather than as a purely bag-of-ngrams or purely sequential task.
3. Joint text-network representation
A major innovation in CleanNews is its embedding strategy, which combines textual information with user network structure. Text embedding uses DeBERTa, while node embedding uses Node2Vec to represent users based on graph structure such as follower or followee relations and propagation trees. For a post authored by user , the fused representation is defined as
where is the DeBERTa embedding for the post content and is the Node2Vec embedding for the user (Cotelin et al., 1 Sep 2025).
The paper characterizes this as 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. In the simplified training pseudocode, word embeddings and node embeddings are computed separately, concatenated, and used for supervised training against ground-truth classes. The stated effect is that the training process becomes sensitive both to what is posted and to who posts it, with network position acting as an additional signal rather than as a post hoc metadata feature (Cotelin et al., 1 Sep 2025).
This fusion is the sense in which the architecture is “network-aware.” It does not merely append graph-derived statistics to a downstream classifier; it incorporates node representations into the same learned input space as the textual representation. The paper further reports, through ablation studies, that Node2Vec network embeddings contribute a measurable performance boost and that DeBERTa consistently outperforms classic embeddings such as GloVe when combined with network features (Cotelin et al., 1 Sep 2025).
4. Mitigation by network immunization
Once fake news is detected, CleanNews attempts to prevent further spread by immunizing strategically important nodes in the network. The two algorithms used are SparseShield and NetShield. Both target nodes whose removal most efficiently disrupts the largest eigenvalue of the adjacency matrix, which is treated as a measure of spreading potential. SparseShield operates on a sparse representation for memory efficiency and is described as better suited to large networks, while NetShield uses a dense matrix representation and is described as potentially faster for smaller but more interconnected graphs (Cotelin et al., 1 Sep 2025).
In SparseShield, the system scores nodes using the dominant eigenvector of the adjacency matrix, maintains a priority queue, and iteratively selects top- nodes to immunize. If a node is already detected as harmful, its score is penalized according to
The paper states that top scoring nodes are then immunized, for example by warning, suspending, or reducing reach. NetShield follows a similar score computation and immunization process but uses dense rather than sparse representations. A Random Solver, which immunizes nodes without regard to structure or influence, is included as a baseline (Cotelin et al., 1 Sep 2025).
A notable detail is the blocking policy. Harmful nodes themselves are not prioritized for immunization; instead, their proximity in the graph causes their neighbors, described as the most likely propagators, to be immunized. The stated rationale is that this selective blocking improves containment while minimizing service disruption. This is an important correction to a common assumption that mitigation must directly target the source node. In CleanNews, containment is formulated as a network intervention problem rather than only a source-removal problem (Cotelin et al., 1 Sep 2025).
5. Experimental configuration and reported performance
The experimental evaluation uses the Twitter15 and Twitter16 datasets, which contain textual content and propagation information. The class labels are true, false, unverified, and non-rumor. The paper reports that classic preprocessing—stopwords, punctuation, lemmatization, and related steps—provided the best results. Hyperparameters were tuned through grid search over network size, depth, dropout, learning rate, epochs, and related settings, and the experiments included ablation studies and cross-validation (Cotelin et al., 1 Sep 2025).
| Dataset | Reported detection result |
|---|---|
| Twitter15 | 0.758 accuracy, 0.726 F1-score |
| Twitter16 | 0.823 accuracy, 0.823 F1-score |
The detection metrics are defined in the standard way:
0
1
The paper specifically reports that the CNN-BiGRU/Node2Vec + DeBERTa configuration delivers the best multi-class classification performance in its study and that CNN-BiGRU outperforms other hybrid models for both detection and precision in identifying fake news (Cotelin et al., 1 Sep 2025).
For mitigation, the experiments use simulation on real propagation subgraphs. The reported outcome is that SparseShield saved up to ~50% of nodes that would otherwise be infected by misinformation, while NetShield was effective but consistently less so than SparseShield, and the Random Solver often saved only 1 node. These results ground the paper’s claim that CleanNews addresses not only classification accuracy but also downstream diffusion control (Cotelin et al., 1 Sep 2025).
6. Position within the news-analysis literature
CleanNews belongs to a broader family of systems that restructure, evaluate, compare, or contextualize news, but its problem formulation is narrower and more interventionist. Earlier work on conceptual news browsing used Formal Concept Analysis to restructure unstructured news into a multi-level tree interface for navigation rather than to detect misinformation or modify diffusion dynamics (Aqle et al., 2018). SciLens, by contrast, evaluates article quality and trustworthiness in real time using content indicators, news-context indicators, social-media context indicators, and domain-expert reviews, but it is framed as quality assessment rather than fake-news containment (Romanou et al., 2020).
NEWSSENSE addresses a different layer of the problem: cross-document evidence and context. It augments a focus article with inline highlights indicating whether specific claims are supported or contradicted by other articles, using reference-free fact verification through article clusters, atomic claim extraction, filtering, and NLI-based claim linking (Milbauer et al., 2023). A bias-aware news recommendation system, in turn, exposes bias scores and recommends same-topic articles from outlets with different political leanings, focusing on reader awareness rather than network immunization (Patankar et al., 2018). NeuS addresses framing bias through neutral multi-news summarization, with a hierarchical model that learns to neutralize news content from title to article, again without any graph-based containment component (Lee et al., 2022).
This situates CleanNews as part of the misinformation branch of news technology rather than the navigation, summarization, or reading-assistance branches. The architecture is closest in spirit to the graph-based social aggregator that replaces item voting with voting on relations between news items and adds a blockchain-based incentive layer, because both systems embed misinformation within an explicitly networked model of interaction (Ziashahabi et al., 2020). Even so, their mechanisms differ: CleanNews uses learned text-plus-network representations and immunization algorithms, whereas the blockchain system relies on relation voting, stake-weighted matrix factorization, and fraud-proof protocols.
7. Data resources, ecosystem analysis, and research implications
CleanNews was evaluated on Twitter15 and Twitter16, but adjacent resources indicate how network-aware fake-news research can be broadened. NELA-GT-2022 provides 1,778,361 articles from 361 outlets, includes outlet-level veracity labels from Media Bias/Fact Check for 337 of 361 sources, and tracks 346,283 distinct tweets embedded within articles (Gruppi et al., 2022). Infini-News processes and enriches over 1.35B Common Crawl news articles, adds three language-detection columns, resolves a country of origin for 83.4% of articles across 222 countries, and provides suffix-array indexes that support arbitrary text-pattern search in sub-second time (Lazzaroni et al., 18 May 2026).
At the level of media-ecosystem structure, NewsCompare offers another relevant perspective by crawling websites, comparing text content, and building hyperlink graphs to identify content similarity, influence networks, and clusters of sites with strong interlinking and similar content (Pop et al., 2019). Islander similarly clusters articles into events, estimates Suspicion Level, Popularity Level, and Sentiment Level, and provides a web interface for browsing multi-source coverage and possible manipulation patterns (Huang et al., 2022).
A plausible implication is that these corpora and ecosystem-mapping systems could complement CleanNews in at least three ways: by expanding training distributions beyond rumor datasets, by enabling longitudinal and cross-national evaluation, and by connecting social-platform mitigation to outlet-level or site-level influence structure. This implication is not claimed directly in the CleanNews paper, but it follows naturally from the coexistence of network-aware mitigation architectures, large-scale news corpora, and systems for cross-source evidence, quality estimation, and influence mapping (Cotelin et al., 1 Sep 2025).