Prediction of cascade structure and outbreaks recurrence in microblogs

Abstract

Cascades are formed as messages diffuse among users. Diffusion process of viral cascades is analogous to spread of virus infection. However, until recently structure properties of viral cascades are quantified and characterized due to available diffusion datasets and increasing knowledge towards it. The virality of structure is a notion for characterizing structural diversity of cascades, but relationship between structural virality and shape of cascades is not highly revealed. We address this problem in a more intuitively way under the help of visualization methods and define a new problem to predict future structure of cascades from the perspective of time. Whether structure of cascades can be predicted, how early it is perceived and which features play key roles in the future propagation structure are discussed in details. Results obtained have a precision rate ranging from 86% in an hour to 97% in a day, indicating future cascade structure can be predicted when proper features are chosen. And the hierarchical tree features which related with structural virality are proven to play an important role in cascade size prediction. Viral cascades often lead to phenomenon of outbreaks recurrence, which has never been discovered formally before. Prediction in outbreak recurrence can also achieve non-trivial performance under the same prediction framework. Moreover, outbreak recurrence is shown to play significant impacts on structure virality of cascades. Our research is especially useful in understanding how viral cascades and events are formed, as well as exploring intrinsic factors in cascade prediction.