Approximate Collapsed Gibbs Sampling with Expectation Propagation

This is Python 3 code for clustering latent variable models with approximate Gibbs sampling.

This repo contains example python code for the paper Approximate Collapsed Gibbs Clustering with Expectation Propagation and Scalable Clustering of Correlated Time Series using Expectation Propagation

Overview

• The experiments folder stores example python scripts for both correlated time series clustering and robust mixture modeling.
• The output folder stores output for the example python scripts.
• The ep_clustering folder stores the python module code for the various Gibbs sampling methods for a variety of models. See ep_clustering/README.md for additional details

Installation

Install add the ep_clustering/ folder to the python path.

Requirements: python 3+, matplotlib, ipython, joblib, numpy, scipy, pandas, scikit-learn, tqdm, seaborn

Usage Example

For correlated time series clustering:

• Walk through experiments/synthetic_timeseries_clustering_example.py for an overview of the API
• Run experiments/synthetic_timeseries_compare_example.py to compare naive, collapsed, and approx-EP Gibbs. (This script takes a while). Below is example output comparing normalized mutual information (NMI) of each sampler's inferred clustering compared to the truth

NMI vs Iteration	NMI vs Time

For robust mixture modeling with the Student's-$t$ distribution:

• Walk through experiments/synthetic_mixture_clustering_example.py for an overview of the API
• Run experiments/synthetic_mixture_compare_example.py to compare naive, blocked, and approx-EP Gibbs.

NMI vs Iteration	NMI vs Time

Release History / Changelog

• 0.1.0
  • The first release (Dec 2019)
  • Over-engineered code from the beginning of my PhD

Meta

Christopher Aicher - [email protected]

Distributed under the MIT license. See LICENSE for more information.