Explaining Social Media Engagement: Interpretable Deep and Machine Learning for Twitter Retweets

Understanding content diffusion on social media is still a wide study challenge. There is no simple answer to the question of why some content is more widely spread than others. To contribute to this, this thesis combines transformer-based, gradient boosting, and NN regression models. These are combined with explainability tools to identify linguistic, topical, and socio-psychological features that drive engagement on Twitter. By combining features based on user-level indicators, sentiment analysis, topic probabilities, and LWIC categories, among others, the thesis aims to provide methods with both accuracy and interpretability. The results obtained aim to not only find which features predict virality but also why they matter and how they are related, allowing a more interpretable understanding of content spread. This is implemented on a wide dataset of climate-change related tweets.