Inferring Latent Co-activation Patterns for Information Diffusion

Different diffusion models have been proposed in previous literature to model information diffusion, in which each node is often assumed to be independently influenced by its parents. More recently, some have begun to challenge this assumption based on the observation that structural and behavioral dependency among the parent nodes exerts a notable role in diffusion within networks. In this paper, we postulate that a node is independently influenced by a set of latent co-activation patterns of its parents, instead of the parents directly. We integrate the latent class model with the conventional independent cascade model where each latent class corresponds to a particular co-activation pattern of the parent nodes. Each parent activation is essentially first "projected" onto the latent space and then "reconstructed" before exerting its influence onto the child nodes. The coactivation patterns are to be inferred based on the information cascades observed without using the connectivity related cues except the information of direct parents. We formulate the co-activation pattern identification problem and the diffusion network inference problem under a unified probabilistic framework. A two-level EM algorithm is derived for inferring the model parameters. We applied the proposed model to a meme dataset and two social network datasets with promising results obtained. Using the results obtained based on the meme dataset, we also illustrate how the identified co-activation patterns can support the analysis of dependency among online news media.