Source-linked AI summary
Fake News Identification on Twitter with Hybrid CNN and RNN Models
Oluwaseun Ajao, Deepayan Bhowmik, Shahrzad Zargari
TL;DR
The paper addresses the growing propagation of fake news on social media by proposing a deep-learning framework for detecting and classifying fake news in Twitter posts. Using CNN and LSTM models, the approach achieves 82% accuracy and identifies relevant features without prior domain knowledge.
Problem
The paper asks whether semantic or linguistic characteristics of fake news on Twitter can be identified automatically without prior knowledge of the domain or news topic.
Method
The framework combines convolutional neural networks with long short-term memory recurrent models for content-based fake-news detection, using a text-only hybrid implementation.
Results
82% accuracy was achieved for fake-news classification on Twitter, while the LSTM-CNN hybrid achieved 74% accuracy and a 39.7% FMeasure.
Takeaways & Limitations
The approach identifies relevant features associated with fake news posts without prior knowledge of the topics being discussed.
Takeaways & Limitations
The deep-learning models were trained on a small dataset of 5,800 tweets, limiting training data available for model robustness.
Abstract
from arXiv · showhide
The problem associated with the propagation of fake news continues to grow at an alarming scale. This trend has generated much interest from politics to academia and industry alike. We propose a framework that detects and classifies fake news messages from Twitter posts using hybrid of convolutional neural networks and long-short term recurrent neural network models. The proposed work using this deep learning approach achieves 82% accuracy. Our approach intuitively identifies relevant features associated with fake news stories without previous knowledge of the domain.
1 INTRODUCTION
Fake-news propagation on social media creates serious societal and financial concerns, motivating automatic identification of deceptive Twitter content. The section asks whether linguistic or semantic features can be identified and truth assessed without prior domain knowledge.
- Fake news affects politics, education, financial markets, and public safety through social-media dissemination.
- A hacked Associated Press account illustrates how a false tweet can reach a large audience and trigger major financial consequences.The account had almost 2 million followers, and the cited false message preceded a 136-billion-dollar loss on the S&P Index in two minutes.
- Fake-news circulation is presented alongside other harmful social-media uses, including propaganda, hate crimes, cyberbullying, and terrorist recruitment.
- The paper asks whether Twitter messages can be classified as authentic or false using content-based semantic or linguistic features.
- The authors frame automatic detection as useful for separating fake messages from authentic ones without prior knowledge of the news topic.
2 RELATED WORKS
Prior Twitter credibility and rumor research uses temporal, structural, linguistic, user, and content-based indicators. The paper positions neural models as a way to combine automated text interpretation with fake-news detection.
- Neural networks have shown high accuracy and precision for text clustering and classification and can detect spatio-temporal trends in social-media propagation.
- The authors describe their hybrid approach as an initial attempt to use neural models for detecting the origin and propagation of fake-news posts.
- Earlier rumor-classification work grouped indicators into temporal, structural, and linguistic features.
- Related studies assessed fake images, phishing, spam, and tweet credibility using user attributes, URLs, networks, keywords, mentions, retweets, and tweet length.
3 METHODOLOGY
The methodology uses deep learning to automatically extract Twitter-post features and classify fake news, with LSTM, dropout, and LSTM-CNN variants evaluated on labeled rumor data.
- The proposed approach automatically identifies Twitter-post features without prior knowledge of the subject domain and classifies fake news using text and images.
- 3.4 Incorporating Convolutional Neural Network: Although the broader proposal mentions text and images, the hybrid implementation reported here uses a text-only approach.
- The study implements three neural variants: plain LSTM, LSTM with dropout regularization, and LSTM combined with a 1D CNN.
- 3.1 The Deep learning Architectures: The LSTM-CNN model places a 1D CNN after word embeddings and adds max pooling to reduce dimensionality while preserving depth.
- The dataset contains approximately 5,800 tweets centered on five rumor stories, labeled as rumors or non-rumors.
- Ten-fold cross-validation and tweet padding were applied to the dataset for model training and processing.
4 EVALUATION, RESULTS AND DISCUSSION
The experiments report 82% accuracy for fake-news classification, while the plain LSTM outperforms the tested dropout and LSTM-CNN variants. The results also indicate that limited training data constrains performance.
- 82% accuracy was achieved for fake-news classification without prior knowledge of the topics being discussed.
- The plain vanilla LSTM achieved the best reported performance across precision, recall, FMeasure, and accuracy.
- The LSTM-CNN hybrid achieved 74% accuracy and a 39.7% FMeasure, outperforming the dropout-regularized model but receiving lower appreciation than plain LSTM.
- The authors attribute weaker neural-network performance to insufficient training examples and expect additional reaction data to improve robustness.
- The PHEME state-of-the-art result had 68% precision, which remained higher than the results obtained in this study.
5 CONCLUSION AND FUTURE WORK
The paper presents a hybrid deep-learning approach for detecting fake news on Twitter, reporting 82% accuracy while requiring relatively few tweets. Future work targets image disambiguation, larger training data, and tracing the origins and locations of fake posts.
- 82% accuracy was achieved for fake-news detection on the PHEME Dataset, beating the reported state of the art.
- The hybrid deep-learning approach determines message veracity without requiring a very large number of tweets about an event.
- Further work examines inferring fake-news locations, origins, and propagating authors, with minimal computational resources as a desired boundary.
- The current dataset contains 5,800 tweets, motivating plans to add hundreds of thousands of user reactions to improve model robustness.