Seeing the forest for the trees: new approaches to forecasting cascades

Patrick Butler, Naren Ramakrishnan

Abstract

Cascades are a popular construct to observe and study information propagation (or diffusion) in social media such as Twitter. and are defined using notions of influence, activity, or discourse commonality (e.g., hashtags). While these notions of cascades lead to different perspectives, primarily cascades are modeled as trees. We argue in this paper an alternative viewpoint of cascades as forests (of trees) which yields a richer vocabulary of features to understand information propagation. We develop a framework to extract forests and analyze their growth by studying their evolution at the tree-level and at the node-level. Moreover, we demonstrate how the structural features of forests, properties of the underlying network, and temporal features of the cascades provide significant predictive value in forecasting the future trajectory of both size and shape of forests. We observe that the forecasting performance increases with observations, that the temporal features are highly indicative of cascade size, and that the features extracted from the underlying connected graph best forecast the shape of the cascade.