gabrieag/bayesian-simplex-clustering

Bayesian clustering with simplicial mixture models

Bayesian simplex clustering

The Bayesian simplex mixture (BSM) model implements simulation and inference algorithms for a simplicial mixture model with component multi-variate t distributions.

Simplicial models are typically used in text-based information retrieval. For instance, the latent Dirichlet allocation (LDA) model allocates topics to documents within a corpus based on their word statistics. In the Bayesian simplex mixture (BSM) model, documents are replaced by sets of continuous data. Each set is assumed to originate from a simplicial mixture of classes with set-specific mixing proportions. Words in the dictionary are replaced by symbols, where each symbol is a mixture component, and classes arise as mixtures of symbols.

The BSM models each component as a heavy-tailed distribution with symbol-specific parameters. Parameters are equipped with conjugate prior distributions as per the Bayesian paradigm. The posterior distributions over both parameters and hidden variables are estimated by an approximate variational inference algorithm.

bayesian-simplex-clustering's Issues

Potential bug caused by class attributes

Asher pointed out that the Dirichlet, Gauss-Gamma and Gauss-Wishart distribution classes in __dist__.py have class attributes, which may lead to subtle bugs.

This can be fixed by making the attributes of param part of the class.

Simplify API

Simplify the API. See tests/py for reference.

Variable renaming

Rename variables and classes for better PEP8 compliance. See segmentation.py for reference.