Andrew Thomas (Department of Statistics, Carnegie Mellon University)
Abstract. A key promise of social networks is the ability to detect and model the correlation of personal attributes along the structure of the network, in either static or dynamic settings. The basis for most of these models, the Markov Random Field on a lattice, has several assumptions that may not be reflected in real network data, namely the assumptions that the process is stationary on the lattice, and that the ties in the model are correctly specified. Additionally, it is less than clear how correlation over longer distances on networks can be adequately specified under the lattice mechanism, given the assumption of a stationary process at work.
Based on concepts from generalized additive models and spatial/geostatistical methods, I introduce a class of models that is more robust to the failure of these assumptions, more flexible to different definitions of network distance, and more generally applicable to large-scale studies of network phenomena. I apply this method to outcomes from two large-scale social network studies to demonstrate its use and versatility.