Predictive updating methods with application to bayesian classification
When comparing models, we’re mainly interested in expressions containing theta, because is a prior, or our belief of what the model parameters might be.
The two most important methods are Monte Carlo sampling and variational inference. The excerpt from The Master Algorithm has more on MCMC.In the spectrum of Bayesian methods, there are two main flavours. The latter contains the so-called nonparametric approaches.Modelling happens when data is scarce and precious and hard to obtain, for example in social sciences and other settings where it is difficult to conduct a large-scale controlled experiment.Latent Dirichlet Allocation is a method that one throws data at and allows it to sort things out (as opposed to manual modelling).
It’s similar to matrix factorization models, especially non-negative MF.If one also takes the prior into account, then it’s maximum a posteriori estimation (MAP). Note that choosing a model can be seen as separate from choosing model (hyper)parameters.