Title: Bayesian fusion
Authors: Hongsheng Dai - University of Essex (United Kingdom) [presenting]
Murray Pollock - University of Warwick (United Kingdom)
Gareth Roberts - University of Warwick (United Kingdom)
Abstract: An exact Bayesian fusion algorithm is presented, which can carry out perfect inferences for the unification of distributed data analysis. The new method uses parallel but coalesced Markov processes to drive distributed Monte Carlo draws to a Monte Carlo sample from the posterior of the full data. The Markov processes are simulated via path-space rejection sampling for diffusion processes. The methodology of this exact Bayesian fusion algorithm explained why existing methods do not provide good results and how to correct approximated draws of existing methods in order to obtain exact samples. Its approximate version, the sequential Bayesian fusion algorithm, can be implemented in parallel for big data analysis.