Hi All
Welcome to Bayesian Business Intelligence my first attempt at blogging. My goal with this blog is to provide some rigorous Bayesian solutions to common business intelligence (BI) problems. My target audience for this blog is mostly my coworkers at in BI GREE, however no matter where you work (some restrictions apply) please feel free to message me with any questions you have.
The blog will focus at least on the short term on Bayesian model selection, in particular applications of the reversible jump Markov Chain Monte Carlo (rjMCMC). My goal is to present both theoretical derivations and sql / r code examples so that the information present here can get past interesting and achieve useful.
Cheers Gavin Steininger
Hi Gavin,
Love what you have done with the blog, good read and there is a definite need for more of this information! For myself, I have been thrown into the world of MCMC since September for my thesis. My background is in physics but my thesis is linked to an engineering master. Besides a Metropolis algorithm, I have applied a single component MCMC and an MCMC scheme that creates proposals from a multivariate normal distribution with covariance to combat the poor mixing I have been experiencing.
When reading up on improvements on single component MCMC, I stumbled on one of your blogs in which you applied principal component analysis. Although my thesis is almost over, I am interested in the possibility of using a principal component analysis of the single component version of the algorithm. I have had some setbacks on the implementation of the PCA-scheme and am hoping I could pick your brain for some advice.
Kind regards,
Djef
LikeLike
Absolutely.
With out really understanding your problem I would give a few suggestions.
1) Parallel tempering / REM is super easy to implement and solves many problems.
2) In MCMC cheating is always a good idea. That is if you know to parameters are very collated because of the physics of the forward model then sample their product/sum and a percentage.
3) I know you are at the end but annealed importance sampling might also be a good approach.
4) if you have derivative information about your problem appendix d in my thesis might be worth trying.
LikeLike