Data Driven Modeling of Continuous Time Information Diffusion in Social Networks

Abstract

Online social networks can detailedly and accurately record the activities of human beings and the trajectories of information dissemination over time, which provides us an opportunity to understand the information diffusion process from a renewed, more realistic, data driven modeling dimensionality. In consideration of two fundamental behaviors (viewing and sharing) involved in information diffusion, we propose a stochastic, heterogeneous, continuous-time delay Unknown-View-Share-Removed (UVSR) model to characterize the information diffusion process. The UVSR model introduces four parameters to describe the diffusion probability and speed: viewing/sharing probability/delay. These parameters are subject to some sort of distributions from the actual data, or based on empirical assumptions. To validate the model, we collect and analyze large number of information cascades (tree structure) diffused in WeChat network. We find that the viewing delay and sharing delay are approximately subject to log-normal and power-law distributions respectively, and the sharing probability follows a Gaussian distribution. Driven by these empirical findings and a constant viewing probability assumption, our model can reproduce numerous key features of information diffusion process in both topology and temporal dynamics, such as cascade size distribution, structural virality, life span distribution and relative propagation speed. Our work contributes to a better understanding of the topological features and temporal dynamics of information diffusion from a continuous time, stochastic modeling view.