Dual Dirichlet Process Gaussian Mixture Model

Code for our paper

Non-parametric clustering over user features and latent behavioral functions with dual-view mixture models
Lumbreras A., Guégan M., Velcin J., Jouve B.
Computational Statistics (2016)

Notes

• DP-GMM: default
• fixed-GMM: set alpha=0 and dont sample it
• single: use sample_z() and do not sample feature view

Multi DP-GMM is a double dirichlet process that simultaneously does:

• Cluster users with respect to their attributes.
• Cluster users with respect to their behaviors.

Inference is done with Monte Carlo methods: Gibbs Sampling and Adaptive Rejection Sampling.

Read the paper for more information.

Some interesting readings

• On the Bogosity of MCMC Diagnostics
• Sampling from Wisharts with scalar degrees of freedom
• The Wishart and Inverse Wishart distributions