Assessing temporal event recurrence as a feature for fake news detection using deep learning

The rapid spread of misinformation through online news platforms and social media has increased the need for effective fake news detection methods. While prior studies have mainly focused on textual, linguistic, and social-contextual features, the temporal nature of reported events remains insufficiently explored. This study investigates whether the recurrence pattern of events can provide useful signals for fake news detection. In particular, news articles were analyzed in terms of recurring and non-recurring events categories, and two deep learning models, BERT and Bi-LSTM, were employed for comparative evaluation. A temporally annotated subset of news articles was constructed and manually labeled using a dual-annotator strategy. Experimental findings show that BERT outperformed Bi-LSTM across all evaluation metrics, achieving stronger accuracy, precision, recall, F1-score, and AUC-ROC. The results further suggest that recurring event structures may offer meaningful contextual cues for distinguishing credible and non-credible news narratives. This work highlights the value of temporal event modeling as a complementary dimension in fake news detection and provides a foundation for future large-scale and socially aware misinformation research.