Title: Efficient marginal likelihood estimation by subsampling thermodynamic integration
Authors: Khue-Dung Dang - University of Melbourne (Australia) [presenting]
Matias Quiroz - University of Technology Sydney (Australia)
Robert Kohn - University of New South Wales (Australia)
Abstract: In Bayesian statistics, the marginal likelihood has a crucial role in model selection. However, it is difficult to compute the marginal likelihood because that requires integrating over all model parameters. Using ideas from thermodynamic integration (TI), the marginal likelihood of a model can be computed via Markov chain Monte Carlo on a series of modified posterior distributions. This method is easy to implement and requires little tuning but can be computationally costly for large data sets. We propose a method to speed up the estimation of the marginal likelihood by combining TI with data subsampling. We apply the new method for modelling binary and time series data sets and show that it is significantly faster than standard TI yet gives similar results.