A0358
Title: Fast MCMC chains for Bayesian posterior inference in graphical models
Authors: Reza Mohammadi - University of Amsterdam (Netherlands)
Ilker Birbil - University of Amsterdam (Netherlands)
Marit Schoonhoven - University of Amsterdam (Netherlands)
Lucas Vogels - University of Amsterdam (Netherlands) [presenting]
Abstract: The purpose is to treat a problem as old as statistics: Discovering the unknown parameters of a distribution. A Bayesian framework is used for this. That means the posterior distribution is set out to be discovered. Markov chain Monte Carlo (MCMC) methods are a popular tool for this. Traditional MCMC strategies, such as reversible jump or birth-death algorithms, are still popular, despite suffering from a slow exploration of the parameter space. An alternative is offered by approximating the detailed balance condition, creating an algorithm that can traverse the entire parameter space in a single iteration. The power of the algorithm is shown in the field of Gaussian graphical models, Ising models, and feature selection, but it is noted that its applicability reaches beyond those examples. In fact, it is applicable to most Bayesian posterior inference.
