DOC create demo for BarsK10V900 MixModel with VB

Make a "case study" IPython notebook that compares many runs of a single learning algorithm on a toy dataset, using a plain-vanilla mixture model.

Here's a (rough) example of what I'd like you to produce https://bitbucket.org/michaelchughes/bnpy-dev/wiki/demos/Demo-HashtagK9

Source notebook: https://bitbucket.org/michaelchughes/bnpy-dev/src/ea3bad/demos/Demo-HashtagK9.ipynb?at=master

This basic template shows some how to use the current bnpy code to make a few plots that can help diagnose issues. You should copy this notebook and use it as a starting point for your experiments.

As a bare minimum, you should probably do a similar analysis that explores how the number of clusters K influences the learned cluster parameters. What happens when K is way too big? Way too small?. Ideally, you will include slightly more thorough experimental analysis (maybe a few more plots) and most importantly, better plain-English discussion of the results.

BarsK10V900 toy data

This "toy bars" dataset is inspired by this famous paper that introduced a Gibbs sampler for topic models. Have a look at datasets/BarsK10V900.py for details.

bnpy.run('BarsK10V900', 'MixModel', 'Mult', 'VB', ...)

BarsK10V900 toy data

Comments (1)