Toward capturing heterogeneity for inferring diffusion networks: A mixed diffusion pattern model

Abstract Inferring diffusion network structure from observed cascades has attracted tremendous attention due to its utmost significance for many applications in online social network (OSN) analysis. Most previous studies assume that information diffuses with a uniform diffusion pattern. However, in OSNs, user interactions usually show different preferences and different speeds, and hence the diffusion processes are heterogeneous and show diverse diffusion patterns. It is difficult for traditional methods to capture the heterogeneity of information diffusion processes in OSNs. In this paper, we study the problem of inferring diffusion networks based on multiple latent diffusion patterns. To this end, we first analyze massive users’ retweeting behaviors to investigate pairwise information transmissions. This analysis allows us to present a reasonable formulation of pattern-based pairwise information transmission probabilities to model the diffusion processes. Then, we incorporate multiple latent diffusion patterns into a probabilistic mixture model to infer diffusion network structures by fitting the observed cascades. We provide the estimation method of our proposed model based on Expectation Maximization (EM) algorithm. The results of experiments conducted on real OSN datasets demonstrate the superior performance of our model in inferring diffusion networks and show that our model can discover latent diffusion patterns effectively.