This package implements three different Bayesian non-parametric methods for inferring the number of clusters in a mixture model, and cluster assignemnts.
- A Gibbs sampler which makes multiple passes over the data (batch).
- A particle filter using the resampling scheme described by Fearnhead (2004) (online).
- A particle filter using the resampling scheme of Chen and Liu (2000) (online).
These can fit variants of a Dirichlet-process mixture model, with a conjugate
prior for the cluster parameters. Currently uni- or multivariate normal
distribution components are supported (via a NormalInverseChisq and
NormalInverseWishart prior). In principle the code could be generalized to
use any of the distributions in
ConjugatePriors.jl but this
requires some additional methods (for doing online updates of the necessary
sufficient statistics and calculating the collapsed/marginal likelihoods).