Parallel Data-Driven Modeling of Information Spread in Social Networks

Models of information spread in social networks are widely used to explore the drivers of content contagion and to predict the effect of new information messages. Most of the existing models (aggregated as SIR-like or network-based as independent cascades) use the assumption of homogeneity of an audience. However, to make a model plausible for a description of real-world processes and to measure the accumulated impact of information on individuals, one needs to personalize the characteristics of users as well as sources of information. In this paper, we propose an approach to data-driven simulation of information spread in social networks which combines a set of different models in a unified framework. It includes a model of a user (including sub-models of reaction and daily activity), a model of message generation by information source and a model of message transfer within a user network. The parameters of models (e.g. for different types of agents) are identified by data from the largest Russian social network vk.com. For this study, we collected the network of users associated with charity community (~33.7 million nodes). To tackle with huge size of networks, we implemented parallel version of modeling framework and tested it on the Lomonosov supercomputer. We identify key parameters of models that may be tuned to reproduce observable behavior and show that our approach allows to simulate aggregated dynamics of reactions to a series of posts as a combination of individual responses.