Multi-Source-Driven Asynchronous Diffusion Model for Video-Sharing in Online Social Networks

Characterizing the video diffusion in online social networks (OSNs) is not only instructive for network traffic engineering, but also provides insights into the information diffusion process. A number of continuous-time diffusion models have been proposed to describe video diffusion under the assumption that the activation latency along social links follows a single parametric distribution. However, such assumption has not been empirically verified. Moreover, a user usually has multiple activated neighbors with different activation times, and it is hard to distinguish the different contributions of these multiple potential sources. To fill this gap, we study the multiple-source-driven asynchronous information diffusion problem based on substantial video diffusion traces. Specifically, we first investigate the latency of information propagation along social links and define the single-source (SS) activation latency for an OSN user. We find that the SS activation latency follows the exponential mixture model. Then we develop an analytical framework which incorporates the temporal factor and the influence of multiple sources to describe the influence propagation process. We show that one's activation probability decreases exponentially with time. We also show that the time shift of the exponential function is only determined by the most recent source (MRS) active user, but the total activation probability is the combination of influence exerted by all active neighbors. Based on these discoveries, we develop a multi-source-driven asynchronous diffusion model (MADM). Using maximum likelihood techniques, we develop an algorithm based on expectation maximization (EM) to learn model parameters, and validate our proposed model with real data. The experimental results show that the MADM obtains better prediction accuracy under various evaluation metrics.