On a Topic-based Information Diffusion Model

Abstract

Information diffusion over a social network can be modeled as stochastic processes of state changes. In this paper, we propose an information diffusion model that takes into account topics of information. More specifically, the proposed model determines the diffusion probability for a directed link by using the content attribute of a target document that will propagate over the link, which represents the topic distribution in the document, and the link attribute that expresses the topic distribution in documents that have propagated over the link. The number of model parameters to be learned is only twice of the number of topics considered, which is much less than the one for traditional models, and they can be efficiently and accurately learned from observed diffusion sequences based on the framework of the maximum likelihood estimation. Through an experiment using real world retweet sequences, we confirmed that the proposed model allows us to estimate the length of an information diffusion sequence more accurately compared to an existing model that do not consider topics at all.