dentifying Inequalities in Information Diffusion on Social Media: A Study of Twitter Cascades Across User, Content, and Contextual Groups

How do user, content, and contextual characteristics influence the growth of cascades on social media? In this study, we address this question by analysing over 1.9 million tweets their corresponding retweets to investigate how user characteristics (e.g., gender, country, verification status, number of followers), content veracity (e.g., true and false), and contextual factors (namely the time of day) influence diffusion dynamics.

We model cascade growth using an analytical function with two parameters (initial diffusion rate and diffusion decay rate and assess their distributions using the Generalised Gamma Distribution (GGD), a flexible family of distributions that can adapt its shape through parameter tuning to represent a wide range of known distributional forms, including heavy-tailed behaviours. Group-based comparisons reveal clear asymmetries: male users, verified accounts, and users with many followers tend to initiate larger cascades, while female and unknown-gender users show statistically equivalent patterns.

Although the size distribution of true claims exhibited heavier tails, indicating the presence of some very large cascades, veracity did not significantly increase the expected number of retweets when controlling for other factors in a regression model. Simulations based on fitted parameters successfully replicated the overall cascade distribution but failed to reproduce differences between groups, likely due to the use of a fixed network structure. Regression models for cascade size confirmed the strong influence of user visibility and verification. However, the models explaining the initial diffusion rate and diffusion decay rate accounted for only a small proportion of their variance. This suggests that the dynamics of initial growth and subsequent slowdown of cascades are less predictable based on user, content, and contextual features alone, and may be influenced by other unobserved variables.

Our findings highlight how differences in exposure, engagement, and social perception can contribute to biases in information diffusion (especially along gender and visibility lines), offering a basis for future studies in algorithmic fairness and platform governance.