studenttmixture 0.0.2.1

Mixture modeling algorithms using the Student's t-distribution

studenttmixture

Mixtures of multivariate Student's t distributions are widely used for clustering data that may contain outliers, but scipy and scikit-learn do not at present offer classes for fitting Student's t mixture models. This package provides classes for:

1. Modeling / clustering a dataset using a finite mixture of multivariate Student's t distributions fit via the EM algorithm. You can select the number of components using either prior knowledge or the information criteria calculated by the model (AIC, BIC).
2. Modeling / clustering a dataset using a mixture of multivariate Student's t distributions fit via the variational mean-field approximation. Depending on the hyperparameters you select, the fitting process may kill off unneeded clusters, so the number of components in this case acts as an upper bound.
3. Modeling / clustering an infinite mixture of Student's t-distributions (i.e. a Dirichlet process). In practice, this model is fitted using some small modifications to the mean-field recipe and has some of the same advantages and limitations.

(1) and (2) are currently available; (3) will be available in version 0.0.3.

Unittests for the package are in the tests folder.

Installation

pip install studenttmixture

Note that starting in version 0.0.2.1, this package contains C extensions and is therefore distributed as a source distribution which is automatically compiled on install. This is a little less convenient but provides a large speed increase so that studenttmixture is typically significantly faster (see docs for a comparison) than scikitlearn's GaussianMixture when fixed_df is used (optimizing the degrees of freedom is slow).

It is unusual but problems with source distribution pip packages that contain C extensions are occasionally observed on Windows, e.g. an error similar to this:

error: Microsoft Visual C++ 14.0 is required.

in the unlikely event you encounter this, I recommend the solution described under this StackOverflow and links.

Finally, if you for whatever reason prefer the pure Python version, install version 0.0.2, i.e.:

pip install studenttmixture==0.0.2

training for mixture models will run slower but no compilation is required.

Usage

• EMStudentMixture
  
• VariationalStudentMixture
  
• Tutorial: Modeling with mixtures
  

Background

• Deriving the mean-field formula
  

Upcoming in future versions

• Planned for version 0.0.3