Speaker: Prof. Jacek Wesolowski (Technical University of Warsaw)
Title: Bayesian decomposable graphical models which are discrete and parametric
Abstract: Discrete graphical models are typically non-parametric with unknowns being cell probabilities in a multiway table. In contrast, continuous graphical models are Gaussian and thus fully parametric, which considerably reduces the number of unknowns. I will propose a negative binomial decomposable model (a related binomial one is also available) with number of parameters reduced to the size of the under- lying graph. I will also introduce a graph-Dirichlet prior and discuss its properties such as conjugacy and strong hyper-Markov features.
This is a joint work with X. Zeng (Univ. Strasbourg, France) and B. Kolodziejek (Warsaw Univ. Tech., Poland)
