Forecasting user engagement and competing cascades in social media diffusion: A Hawkes-Transformer approach

Abstract

Social media has evolved into a socio-technical infrastructure that shapes public attention, social interaction, and information governance. Understanding how user engagement behaviors, such as retweets, comments, and likes, collectively influence information diffusion is important for forecasting digital dynamics. Using large-scale data from Sina Weibo, this study develops a hybrid Hawkes–Transformer framework that combines the interpretability of self-exciting point processes with the predictive capacity of deep learning. The model captures both interactions within a post and competition across parallel posts within the same trending topic. Empirical results show that retweets strongly amplify diffusion through self-excitation, while comments can suppress diffusion by diverting user attention. In addition, parallel cascades tend to fragment rather than reinforce information flow. By incorporating Hawkes-estimated parameters as structured inputs into a Transformer model, the proposed approach improves predictive performance while retaining interpretability. These findings provide insights into how attention is distributed and competed for in social media environments, with implications for understanding algorithmic visibility and managing information diffusion.