Generalized network vector autoregression

Abstract

With the rapid development of social network platforms, the time series of network data is becoming increasingly available. To model the dynamic user behaviors, a network vector autoregression model is developed, which targets at the continuous type responses. In practice, the discrete type of data (e.g., numbers of posts, user decisions) are frequently collected from the network users. To model such a type of data, we propose a generalized network vector autoregression model in this work. It assumes that a latent continuous variable exists for each node at each time point, which determines the observed response variable. The dynamic and network dependence is assumed based on the latent variables (states). To estimate and make a valid inference of the model, an MCMC (Markov chain Monte Carlo) algorithm is designed and verified by extensive numerical studies.Two real data examples are presented using datasets from a social network platform for illustration.