2015 QPRC

Fox News Network Data Analysis: Bayesian Dynamic Modeling

Karou Irie, Duke University

We propose a Bayesian approach to analyze data on Internet traffic flow among Fox News websites. The observations are time-varying counts (non-negative integers), so the straightforward application of existing state-space models is not available. It is a Big Data problem, with many different types of articles, raising scalability issues; however, sparsity can be exploited in both modeling and computation. These features of the data motivate use of dynamic versions of count data models (Poisson-Gamma models and Multivariate-Dirichlet models), and lead to fitting an interpretable Gravity model that is an equivalent to two-way ANOVA.  The conjugacy of this model enables use of Forward Filtering and Backward Sampling to obtain the posterior distributions. In addition, the Gravity model reveals the underlying structure of traffic networks across websites, allowing the detection of significant flows and flow dynamics, and enabling computational advertisers to better target their ad campaigns.