Abstract:
|
Mixtures of random graphs - i.e., graph mixtures - are families of distributions on graph sets that are in many ways analogous to mixtures of random variables in other statistical contexts. Just as many familiar heavy-tailed distributions can be written as mixtures of more basic light-tailed distributions, so too do continuous mixtures of basic graph distributions lead to ``heavy-tailed'' distributions on graph sets. These families have many useful applications in social network analysis, ranging from modeling of density and reciprocity overdispersion in multi-network exponential family random graph (ERG) models to robust priors for network inference from error-prone informant reports. In this talk I discuss some simple graph mixture families, and show how they can be employed to address a number of current problems in the statistical modeling of social networks.
|