ETD

• Continual Learning
• Deep Learning
• Fake News Detection
• Machine Learning

This thesis studies Fake News Detection in a Continual Learning setting, addressing the limitations of static models in dynamic news environments. A wide range of content-based models, from traditional machine learning approaches to DNNs and Transformer-based architectures, is evaluated on 12 heterogeneous datasets under topic- and time-incremental scenarios. The thesis aims to analyze models and continual learning (CL) techniques for lifelong fake news detection by investigating the impact of catastrophic forgetting in naive sequential fine-tuning compared to the effectiveness and trade-offs of different CL strategies (replay, regularization, and hybrid approaches) in mitigating performance degradation, and the robustness of various model architectures.