I didn’t know this paper by Kevin Murphy (I like his book “machine learning a probabilistic perspective”). It seems very interesting. The derivation I proposed was mainly done to let me better understand some concepts. In this version, I use full covariance matrices and I derive closed form expressions for the update of the parameters. But this is not a finished work. I’m thinking of a follow-up using the Choleski factors of inverse covariances. Then, there will be no need to inverse covariance matrices and it should be much faster. (Other ideas I want to experiment with: combining this approach with an ensemble of random projections to work with high dimensional datasets) I will also play with real world datasets.
Thanks, Pierre-Edouard > On 12 Jan 2018, at 15:23, 'Jon Hough' via Programming > <[email protected]> wrote: > > Thanks, I was working on something similar, and this is very interesting, and > much nicer written than mine. Although I haven't read all of it yet (it is > quite dense), just comparing briefly to the Scikit-Learn implementation, > (see: > https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/mixture/gaussian_mixture.py > ) > they use several possible covariance types > spherical, full, diagonal, tied, > I think based on this paper: https://www.cs.ubc.ca/~murphyk/Papers/learncg.pdf > It seems your implementation is using only spherical type. > > Also, I am wondering why you tested on simulated data - taken from normal > distributions. Have you tried testing on real world datasets? > > Thanks, > Jon > -------------------------------------------- > On Fri, 1/12/18, Pierre-Edouard PORTIER <[email protected]> > wrote: > > Subject: [Jprogramming] JGMM, Mixture Model in J > To: [email protected] > Date: Friday, January 12, 2018, 10:10 PM > > Mixture models in J: > > http://peportier.me/blog/201801_JGMM/ <http://peportier.me/blog/201801_JGMM/> > > Enjoy your weekend, > > Pierre-Edouard > ---------------------------------------------------------------------- > For information about J forums see http://www.jsoftware.com/forums.htm > ---------------------------------------------------------------------- > For information about J forums see http://www.jsoftware.com/forums.htm ---------------------------------------------------------------------- For information about J forums see http://www.jsoftware.com/forums.htm
