DiagGmm: Diagonal Gaussian Mixture Model Prior description class

In comeetie/greed: Clustering and Model Selection with the Integrated Classification Likelihood

Diagonal Gaussian Mixture Model Prior description class

Description

An S4 class to represent a multivariate diagonal Gaussian mixture model. The model corresponds to the following generative model:

π \sim Dirichlet(α)

Z_i \sim \mathcal{M}(1,π)

λ_k^{(d)} \sim \mathcal{G}(κ,β)

μ_k^{(d)} \sim \mathcal{N}(μ,(τ λ_k)^{-1})

X_{i.}|Z_{ik}=1 \sim \mathcal{N}(μ_k,λ_{k}^{-1})

with \mathcal{G}(κ,β) the Gamma distribution with shape parameter κ and rate parameter β. These classes mainly store the prior parameters value (α,τ,κβ,μ) of this generative model. The DiagGmm-class must be used when fitting a simple Diagonal Gaussian Mixture Model whereas the DiagGmmPrior-class must be sued when fitting a CombinedModels-class.

Usage

DiagGmmPrior(tau = 0.01, kappa = 1, beta = NaN, mu = NaN)

DiagGmm(alpha = 1, tau = 0.01, kappa = 1, beta = NaN, mu = NaN)

Arguments

tau	Prior parameter (inverse variance), (default 0.01)
kappa	Prior parameter (gamma shape), (default to 1)
beta	Prior parameter (gamma rate), (default to NaN, in this case beta will be estimated from data as 0.1 time the mean of X columns variances)
mu	Prior for the means (vector of size D), (default to NaN, in this case mu will be estimated from data as the mean of X)
alpha	Dirichlet prior parameter over the cluster proportions (default to 1)

Value

a DiagGmmPrior-class object

a DiagGmm-class object

References

Bertoletti, Marco & Friel, Nial & Rastelli, Riccardo. (2014). Choosing the number of clusters in a finite mixture model using an exact Integrated Completed Likelihood criterion. METRON. 73. 10.1007/s40300-015-0064-5. #'

See Also

DiagGmmFit-class, DiagGmmPath-class

Other DlvmModels: CombinedModels, DcLbm, DcSbm, DlvmPrior-class, Gmm, Lca, MoM, MoR, MultSbm, Sbm, greed()

Examples

DiagGmmPrior()
DiagGmmPrior(tau = 0.1)
DiagGmm()
DiagGmm(tau = 0.1)

comeetie/greed documentation built on Oct. 10, 2022, 5:37 p.m.