Skip to content

Latest commit

 

History

History
110 lines (77 loc) · 3.74 KB

mixture.md

File metadata and controls

110 lines (77 loc) · 3.74 KB

Mixture Models

A mixture model is a probabilistic distribution that combines a set of components to represent the overall distribution. Generally, the probability density/mass function is given by a convex combination of the pdf/pmf of individual components, as

$$f_{mix}(x; \Theta, \pi) = \sum_{k=1}^K \pi_k f(x; \theta_k)$$

A mixture model is characterized by a set of component parameters \Theta=\{\theta_1, \ldots, \theta_K\} and a prior distribution \pi over these components.

Type Hierarchy

This package introduces a type MixtureModel, defined as follows, to represent a mixture model:

abstract type AbstractMixtureModel{VF<:VariateForm,VS<:ValueSupport} <: Distribution{VF, VS} end

struct MixtureModel{VF<:VariateForm,VS<:ValueSupport,Component<:Distribution} <: AbstractMixtureModel{VF,VS}
    components::Vector{Component}
    prior::Categorical
end

const UnivariateMixture    = AbstractMixtureModel{Univariate}
const MultivariateMixture  = AbstractMixtureModel{Multivariate}

Remarks:

  • We introduce AbstractMixtureModel as a base type, which allows one to define a mixture model with different internal implementations, while still being able to leverage the common methods defined for AbstractMixtureModel.
AbstractMixtureModel

  • The MixtureModel is a parametric type, with three type parameters:

    • VF: the variate form, which can be Univariate, Multivariate, or Matrixvariate.
    • VS: the value support, which can be Continuous or Discrete.
    • Component: the type of component distributions, e.g. Normal.

  • We define two aliases: UnivariateMixture and MultivariateMixture.

With such a type system, the type for a mixture of univariate normal distributions can be written as

MixtureModel{Univariate,Continuous,Normal}

Constructors

MixtureModel

Examples

# constructs a mixture of three normal distributions,
# with prior probabilities [0.2, 0.5, 0.3]
MixtureModel(Normal[
   Normal(-2.0, 1.2),
   Normal(0.0, 1.0),
   Normal(3.0, 2.5)], [0.2, 0.5, 0.3])

# if the components share the same prior, the prior vector can be omitted
MixtureModel(Normal[
   Normal(-2.0, 1.2),
   Normal(0.0, 1.0),
   Normal(3.0, 2.5)])

# Since all components have the same type, we can use a simplified syntax
MixtureModel(Normal, [(-2.0, 1.2), (0.0, 1.0), (3.0, 2.5)], [0.2, 0.5, 0.3])

# Again, one can omit the prior vector when all components share the same prior
MixtureModel(Normal, [(-2.0, 1.2), (0.0, 1.0), (3.0, 2.5)])

# The following example shows how one can make a Gaussian mixture
# where all components share the same unit variance
MixtureModel(map(u -> Normal(u, 1.0), [-2.0, 0.0, 3.0]))

Common Interface

All subtypes of AbstractMixtureModel (obviously including MixtureModel) provide the following two methods:

components(::AbstractMixtureModel)
probs(::AbstractMixtureModel)
Distributions.component_type(::AbstractMixtureModel)

In addition, for all subtypes of UnivariateMixture and MultivariateMixture, the following generic methods are provided:

mean(::AbstractMixtureModel)
var(::UnivariateMixture)
length(::MultivariateMixture)
pdf(::AbstractMixtureModel, ::Any)
logpdf(::AbstractMixtureModel, ::Any)
gradlogpdf(::AbstractMixtureModel, ::Any)
rand(::AbstractMixtureModel)
rand!(::AbstractMixtureModel, ::AbstractArray)

Estimation

There are several methods for the estimation of mixture models from data, and this problem remains an open research topic. This package does not provide facilities for estimating mixture models. One can resort to other packages, e.g. GaussianMixtures.jl, for this purpose.