A Review on Fake News Detection and Personalized Recommendation on Social Media using BERT

Abstract

Therapidspreadofmisinformationonsocialmedia platforms poses a serious threat to public trust, social stability, and digital well-being, making fake news detection a critical researchproblem.Toaddressthischallenge,theintegration of intelligent detection models with privacy-preserving mecha- nisms has become increasingly important. This paper presents a comprehensive review of transformer-based fake news detection approaches, particularly those utilizing Bidirectional Encoder Representations from Transformers (BERT), along with Feder- ated Learning techniques for secure and decentralized person- alization. The study analyzes state-of-the-art methods, bench- markdatasets,andkeyperformancemetrics,demonstratinghow BERT-basedmodelseffectivelycapturedeepcontextualsemantics to accurately distinguish fake content from factual information. Inparallel,federatedlearningenablesdistributedmodeltraining while preserving user privacy, making it suitable for deployment in recommendation systems on social media platforms. Further- more,thisreviewdiscussesahybridarchitecturalperspectivethat combines a centralized BERT-based classifier with a lightweight, device-side federated recommender to deliver trustworthy and personalized news feeds. Finally, major challenges such as data imbalance, model explainability, and adversarial manipulation areexamined,andfutureresearchdirectionsareoutlinedtoward robust, interpretable, and ethically aligned artificial intelligence systems for combating misinformation.

Introduction

The rapid spread of information on social media has increased public connectivity but also intensified the spread of misinformation and fake news, which can distort public opinion and undermine trust. Traditional moderation and manual fact-checking methods are too slow and resource-intensive to handle the scale and speed of modern social platforms, creating the need for automated detection systems.

Early fake news detection methods relied on handcrafted textual features with classical machine-learning models, but these approaches lacked robustness and adaptability. The field has since shifted toward deep learning, particularly transformer-based models such as BERT, which capture rich contextual and semantic information and significantly outperform traditional methods. Variants such as blended and dual BERT architectures further enhance stability and generalization.

To improve robustness, researchers have introduced ensemble learning, graph-based methods (GNNs) that model user–content interactions and propagation patterns, and geometric learning techniques. Multimodal approaches that combine text, images, metadata, and external knowledge have also gained importance, as misinformation increasingly spans multiple media types. Knowledge-based and vagueness-aware models improve interpretability and factual grounding, while unsupervised and self-supervised methods address the scarcity of labeled data.

Recent work highlights the importance of privacy, personalization, and human-centric factors. Federated learning enables privacy-preserving, scalable personalization, while studies show that media literacy and critical thinking influence misinformation susceptibility. Despite significant progress, most existing systems focus on only a subset of these dimensions.

Conclusion

This survey highlights the rapid evolution of fake news detection research, tracing its progression from traditional feature-engineered approaches to advanced deep learning, multimodal, and graph-based architectures. Early studies pri- marily employed classical machine learning algorithms suchas Support Vector Machines, Decision Trees, and Random Forests, relying heavily on handcrafted linguistic and stylistic features.Althoughthesemethodsachievedinitialsuccess,their limitedcontextualunderstandingand poorcross-domaingen-eralization motivated the shift toward representation learning techniques. The introduction of transformer-based models, particularly BERT and its variants, marked a significant advancement in fakenewsdetectionbyenablingdeepercontextualandseman- ticmodeling.WorkssuchasAyyubetal.[6]andFarokhianand Rafe[7] demonstratethat contextualembeddingssubstantially improve robustness and adaptability when dealing with noisy and informal social media text. Ensemble learning strategies and hybrid pipelines further enhance generalization and re- silience against domain shifts. Recent research also emphasizes the importance of incor- porating multiple data modalities and relational structures. Multimodalframeworks[10],[11],[17],[20]integratetextual, visual, and contextual information to capture cross-modal inconsistencies and correlations, while graph neural networks and geometric deep learning approaches [3], [8] exploit social interactionandpropagationpatternstodetectcoordinatedmis- information campaigns. These findings indicate that effective misinformation detection requires understanding both content and its social dissemination dynamics. In parallel, interpretability and factual grounding have emergedaskeypriorities.Knowledge-basedmethodsleverag- ing external knowledge graphs [15], [18], along with vague- ness detection techniques [12], enhance transparency and reliabilitybyaligningmodeldecisionswithverifiablefactsand explainable linguistic cues. Additionally, self-supervised and contrastive learning approaches [19] address label scarcity by enabling models to learn discriminative representations from unlabeled data streams. Human-centered and ethical considerations further enrich this research landscape. Studies on media literacy and user behavior [9] underline that misinformation mitigation is not solely a technical challenge but also a socio-cognitive one. Misinformation-aware recommender systems and privacy- preserving personalization frameworks [16], [17] reflect a growingefforttobalancedetectionaccuracy,userengagement, and ethical data usage. Overall, the surveyed literature reveals a clear trajectory toward integrated solutions. Future fake news detection sys- temsmustunifydeepcontextualmodeling,multimodalfusion, graph-based reasoning, and privacy-aware personalization to deliver scalable, interpretable, and user-centric solutions. The field is steadily advancing toward comprehensive frameworks that balance accuracy, transparency, and ethical responsibility in combating misinformation.

References

[1] T. H. C. Chiang, C. S. Liao, and W. C. Wang, “Investigating thedifference of fake news source credibility recognition between ANNand BERT,” *Journal of Information Technology and Media Studies*,vol. 9, no. 2, pp. 45–52, 2022. [2] M. Al-shaqi, D. B. Rawat, and C. Liu, “Ensemble techniques for robustfake news detection,” in *Proceedings of the International Conferenceon Machine Learning and Applications (ICMLA)*, pp. 1–8, 2024. [3] Y. Han, S. Karunasekeran, and C. Leckie, “Graph neural networks withcontinual learning for fake news detection from social media,” *IEEETransactions on Knowledge and Data Engineering*, vol. 32, no. 10, pp.2046–2059, 2020. [4] G. Guler and S. Demirci, “Deep learning based fake news detection onsocialmedia,”InternationalJournalofComputerApplications,vol.180,no. 34, pp. 1–7, 2023. [5] D.Singh,H.Panwar,andS.Chaturvedi,“Designingfakenewsdetectionsystem using machine learning techniques,” International Journal ofAdvanced Research in Computer Science, vol. 16, no. 5, pp. 102–108,2025. [6] S. Ayyub, A. Khan, and R. Malik, “BERT-based blended approach forfake news detection,” Journal of Computational Linguistics Research,vol. 15, no. 3, pp. 215–223, 2021. [7] M.FarokhianandV.Rafe,“FakenewsdetectionusingdualBERTdeepneuralnetworks,”NeuralComputingandApplications,vol.35,no.9, [8] pp.7511–7520,2023. [9] L. Zhang, Y. Liu, and P. Chen, “Fake news detection on social mediausing geometric deep learning,” Expert Systems with Applications, vol.237, pp. 121–130, 2024. [10] N. Alves and M. Costa, “Fake news detection on social media: Thepredictive role of university students’ critical thinking dispositions andnew media literacy,” Computers in Human Behavior, vol. 136, 2022. [11] S. Patel, A. Nair, and J. Lee, “CroMe: Multimodal fake news detectionusing cross-modal tri-transformer and metric learning,” IEEE Access,vol. 11, pp. 134056–134069, 2023. [12] R. Kumar and V. Menon, “Multimodal fake news detection with con-trastivelearningandoptimaltransport,”InformationFusion,vol.103, [13] pp.18–29,2024. [14] H. Li and X. Zhao, “Combining vagueness detection with deep learningto identify fake news,” Journal of Artificial Intelligence Research, vol.76, pp. 310–326, 2023. [15] P.ReddyandM.Thomas,“Acomparativestudyofmachinelearninganddeep learning techniques for fake news detection,” Procedia ComputerScience, vol. 197, pp. 512–520, 2022. [16] D. Bhattacharya and S. Kumar, “Detection of fake news using machinelearning and natural language processing algorithms,” Applied Intelli-gence, vol. 54, pp. 4205–4219, 2024. [17] E.GarzaandM.Rodr´?guez,“Combiningmachinelearningwithknowl-edge engineering to detect fake news in social networks—A survey,”Knowledge-Based Systems, vol. 233, pp. 107–125, 2021. [18] A. Fernandez, S. Iqbal, and R. Green, “Advancing MisinformationAwareness in Recommender Systems for Social Media InformationIntegrity,” Information Processing and Management, vol. 62, no. 4,pp.101–115, 2025. [19] L. Fu, J. Zhang, H. Li, and X. Wang, “Multimodal FakeNews DetectionIncorporating External Knowledge and User Interaction Feature,”WileyOnline Library, 2023. [20] T.-V. La, M.-H. Nguyen, and M.-S. Dao, l fake news detection,” arXivpreprint,2025. [21] X. Chen, L. Zhou, and Y. Li, “An unsupervised fake news detectionframeworkbasedonstructuralcontrastivelearning,”Cybersecurity,vol.8, no. 1, 2025. [22] Y. Zhu, Y. Wang, and Z. Yu, “Multimodal fake news detection: MFNDdataset and shallow-deep multitask learning,” in Proceedings of theInternational Joint Conference on Artificial Intelligence (IJCAI), pp.891–899, 2025.

Copyright

Copyright © 2025 Adithya A A , Hana Habeeb Rahman, Goutham Krishnan, Sylasree V A, Reshma P D. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.