Recent developments in mixed membership modeling
State of the art talk by Elena Erosheva, Department of Statistics, University of Washington, USA
Chair: Irini Moustaki, Wednesday 22nd July, 14.30 - 15.15, Palmeston lecture Theatre, Fisher Building.
Mixed membership models is a class of models for analyzing multivariate data such as attributes collected on individuals. Variants of mixed membership models have recently gained popularity in many fields ranging from genetics to computer science and studies of social networks. A distinctive feature of mixed membership models is the assumption that individuals may combine attributes from several basis categories in a stochastic manner, according to individual proportions of membership in each category. Formulating a mixed membership model requires specifying assumptions at the population, individual, and latent variable levels, and choosing a sampling scheme for generating individual attributes. One could use variations in these assumptions to come up with different mixed membership models.
In this talk, I will review the basic structure of mixed membership models and relate it to other multivariate analysis methods. I will comment on employing likelihood-based, MCMC and variational estimation methods, pointing out their respective advantages and disadvantages. Finally, I will provide an overview of several recent applications of mixed membership models and their extensions.